1 Introduction

Food acts as a prime source of energy for the living organisms that help them to grow and survive on earth. Quality consumption of food plays a crucial role in the absorption of necessary nutrients for complete growth and development of the body. It has become mandatory to inspect the quality of food and ensure the safety of consumers all over the world. Food adulteration or food fraud is the economically inspired means to harm public health. It is the despicable and facile method of collecting huge fortunes that are a great threat to the lives of people. To quench the thirst of greed, people add adulterants to the food products to get the maximum monetary funds by selling the low-valued products at higher prices. Hence, food adulteration has become a profitable business nowadays. The shopkeepers, vendors and other dealers are playing with the lives of common people just for their illegitimate profits. Almost all food being reported is subject to food adulteration including dairy products, grains, seafood, oils, alcoholic drinks, honey, etc. Also, the fruits and vegetables being sold in the market are not as pure as they are injected with harmful chemicals and pesticides. There are multiple ways of impairing the quality of the food. The food is considered to be adulterated if the valuable constituents of the food are removed or if the poor quality of food products are concealed with an actual food product. Sometimes even the alternative food product is replaced by the actual food item. The food quality deteriorates if some unidentified substance is added to the food or its container is made up of any toxic material. Sometimes even an unsafe pesticide or some other chemical substance is contained in it. Food adulteration is not the only means of debasing the food quality. Contamination of food may happen as a natural consequence of a process. For example, microorganisms may deteriorate the fruits and vegetables. Even the spoilage of dairy products, perishable beverages and other food items can be considered as food contamination.

It seems that traditional food safety methods are not enough to control the issue. Therefore, innovative and advanced ways have to be developed that may be used by the common public or less trained people in this field to keep a check on the quality. The methods should be user-friendly and affordable tools should be designed to evaluate the food quality and achieve the desired aim.

1.1 Different Types of Adulterants Added in Food

It is quite obvious that food adulteration is done in such a way that it is not easy to identify the adulterant by the techniques that are available in public laboratories.

1.1.1 Colour as Food Additives

The colour of food is believed to influence the approach to detect food quality. Therefore, scientific research has been redirected to preserving food colour [1]. For the same reason, prohibited colouring agents are common food adulterants even though they cause many types of health hazards. Also, the permissible food dyes are used in large quantities to attract customers [2, 3]. There are other colouring agents like Metanil Yellow and Rhodamine B etc. that are widely used in confectionery products, dried fruits, wines, bitter sodas, juices, sauces, pastes, and spices. Food Dyes including Allura red and sunset yellow are used in Strawberry jelly and wine [4].

1.1.2 Preservatives as Food Additives

Every civilized society is using food preservatives but such practice can pose a threat to public health [5]. Secure and effective preservative production for perishable food products is a subject of intense study. For example, a suitable mixture of potassium lactate and sodium diacetate is observed as an acceptable preservative under refrigeration conditions [6]. Salt is found to be an effective meat preservative but it can cause hypertension [7]. Safety and efficiency of preservatives are the fundamental criteria that have to be considered for long-term food preservation. However, malpractice like the addition of harmful preservatives to food is often reported. One should have provision to control the addition of harmful food preservatives but in the modern world, it’s not possible to ban safe and effective food preservatives. Therefore, appropriate methodologies have to be set up for the proper screening of food preservatives and food quality.

1.2 Need for Artificial Intelligence for Food Adulteration Detection

Artificial Intelligence (AI) can be used as an opportunity in the food industry. It has a major role to support our food system as it can help in precision farming and many other applications in food production and food consumption. It can also be used as a quality control measure in the food sector. AI is changing the way one thinks about food production, quality, delivery etc. and the era of intelligent mobile apps has a big contribution to this transition. The artificial brains can be used efficiently to create food databases and analyze them. It has the potential to create a healthier and more affordable food industry for workers as well as consumers.

The methods used in industries to detect food adulteration are quite expensive and complex. The specialized infrastructure is required for quality evaluation methods. These methods demand intensive manual labour and are sometimes quite tedious and inefficient. Hence, AI can be used as a platform to develop a low-cost automated system that could be used by the end-user to detect adulteration in fruits, vegetables and dairy products.

The different modern methods in this area like electronic tongues, electronic noses [8], computer vision [9], spectroscopy and spectral imaging [10], and so on, have been widely used to detect food quality. These techniques can acquire a large amount of digital data related to food composition and its properties but it is highly important to analyze this data and extract useful information out of it. But it’s a challenging task to bring these methods into real world applications.

AI based techniques shall be used to evaluate the quality and analyze the data. There are many such methods to deal with a large amount of data such as partial least squares [11], artificial neural network (ANN), support vector machine (SVM) [12], random forest [13], k-nearest neighbour (KNN) [14], and so on. For feature extraction, principal component analysis (PCA) [15], wavelet transform (WT) [16], independent component correlation algorithm (ICA) [17], scale-invariant feature transform [18], histogram of oriented gradient [19], and so on. These methods are highly useful in dealing with this type of data. Thus analyzing the current scenario of degradation in food quality has encouraged the researchers to explore the research in this area.

1.3 Motivation for Research

Some of the key points that inspired us to carry out the analysis are as follow.

  • Food adulteration is a global issue. In recent times, the cases of food quality degradation by various unfair means are rising exponentially. Some sort of awareness in this context may prevent life-threatening diseases and save the lives of innocent people who become the victim of this menace.

  • We were keen to know about the data sources and the datasets available for food quality detection.

  • We observed the necessity to gather information about various techniques to detect food adulteration.

  • We discerned to figure out the factors considered to evaluate the food quality and various food products that have been studied to detect adulteration.

  • We wished to explore different deep learning and machine learning applications in the food domain.

  • We desired to study the various existing systems to detect food adulteration and to evaluate food quality.

  • We recognized the need for a systematic review after observing the ongoing research in this area of food adulteration detection. Consequently, the available research is summarized in this study based on extensive and methodical research.

1.4 Our Contributions

Our contribution to carry out this review is summarized as follows.

  • The relevant research articles have been studied by following a systematic review technique. The papers on food quality detection methods have been classified year wise as well as these were categorized based upon their source of extraction such as conference and journal papers.

  • A detailed analysis to study various methods for the detection of food adulteration has been conducted.

  • The online available datasets for different food categories are discussed for research in this field.

  • This survey discusses the various machine learning as well as deep learning applications in the food domain.

  • The existing system is studied that is used to detect food adulteration and to evaluate food quality.

  • The research work done on different food products by researchers is also presented.

  • The last section discusses future research guidelines in the area of food adulteration detection.

1.5 Related Surveys

The surveys conducted earlier have been productive but they cover a different aspect of food adulteration detection. They have not worked on the survey presenting the machine learning and deep learning techniques to detect food adulteration and also the guidelines to conduct a systematic survey are not followed. Bansal et al. [20] have discussed different types of food adulteration done in various food items, health risks imposed by the adulterants and detection methods available for them. They have studied molecular methods to detect biological adulterants in food [20]. Banerjee et al. [21] have aimed to review the available methods of detection of food fraud to focus on the detection of common adulterants and recent advances [21]. Bhargava et al. (2018) have presented a detailed overview of various methods that address the evaluation of the quality of the fruits and vegetables based on their colour, texture, size, shape and defects. These methods include pre-processing, segmentation, feature extraction, and classification. A critical comparison has been carried out based on different algorithms proposed by the researchers to inspect the quality of fruits and vegetables [22]. Zhou et al. [23] have discussed the use of deep learning in the food domain that includes food recognition, calories estimation, quality detection of fruits, vegetables, meat and aquatic products, food supply chain, and food contamination [23]. The popular architectures of deep neural networks are discussed and it has been found that deep learning can be used as a data analysis tool to solve the challenges and problems of the food category.

1.6 Article Organization

The paper in the following sections is structured as follows. Section 2 covers the review methodology used to conduct the survey. The extraction outcomes in the form of resources of publications are discussed in Sect. 3. The preliminaries of food quality detection are discussed in Sect. 4. Section 5 gives a brief description of the techniques to detect food adulteration. Section 6 provides machine learning and deep learning applications in the food domain. Section 7 deals with the findings identified in the survey and the conclusion of the paper and future directions of research are presented in Sect. 8.

2 Review Methodology Followed

The review of food adulteration detection using machine learning and deep learning methods has been conducted using the following steps.

2.1 Development of Review Protocol

A review methodology is followed to conduct a systematic review. Known electronic databases and the topmost conferences related to the research areas have been consulted for conducting the review. After this, the count of selected studies has been narrowed down by the following the inclusion and exclusion principle. Then the research questions have been framed to select the final research studies and the results are collated after following a thorough analysis.

3 Research Questions

The systematic literature review described in this paper relies on a detailed literature survey that has been conducted to study the various approaches followed by researchers to detect food adulteration. For the effective conduct of the systematic review, the following questions have been framed as listed in Table 1

Table 1 Research questions for systematic literature review
Full size table

3.1 Sources of Information

A suitable collection of electronic databases was considered before starting the search process to discern only the relevant research papers. The databases like Google Scholar (www.scholar.google.co.in/), Science Direct (www.sciencedirect.com), IEEE Xplore (www.ieeexplore.ieee.org) and Taylor & Francis (https://www.tandfonline.com/) have been selected to analyze the research studies. Most of the papers have been presented in top food safety and machine learning conferences, and virtually all papers are also covered by Google Scholar. Before the final selection of research articles, the redundant articles on Science Direct and IEEE Xplore have been eliminated.

3.2 Inclusion and Exclusion Criteria

A systematic keyword-based search was carried out to retrieve the pertinent research studies from the electronic databases as presented in Table 2.

Table 2 Keyword based advanced search from the year (2009–2019)
Full size table

This study involves both quantitative and qualitative research articles from the last decade to make sure that the analysis is complete as an effort to work on food adulteration detection. The keywords “food adulteration” and “food quality” led to a large number of results when these were used to find the research articles because this field is explored in different aspects. The abstracts and the titles of the articles have been searched using the search string “Food adulteration/quality detection [with, using, by] [Technique_used].”

The different research studies use multiple ways of writing the same title. Hence, all these options have been considered while searching the related studies so that all of the research articles can be included. The research studies from different journals, conferences, PhD and Masters Thesis have been considered by following an exclusion principle at various stages as shown in Fig. 1.

Fig. 1
figure 1

Review Technique followed

Full size image

Our search came out with 300 research studies as depicted in Fig. 1. which were then filtered based upon their titles and reduced to 210. These studies were further sorted based on their abstract and came down to 150.This number went down to 112 based on full-text. After that, these 112 research articles were studied thoroughly to select a final list of research articles.

4 Extraction Outcomes

The main objective of this review is to identify the available research on food adulteration detection using artificial intelligence and is stated in the form of research questions in Table 1. To address the research question RQ1, the yearly status of research studies about food adulteration detection methods using artificial intelligence has been shown in Fig. 2a. The origin of sources of their publications is depicted in Fig. 2b.Currently, the field of food quality detection is a hot and challenging research area. Hence, the year-wise status of publications from the last decade is presented in Fig. 2a. It is quite evident from the graph that research in this area is continuously growing from the last couple of years. During the thorough analysis, it has also been seen that most of the research articles on food adulteration detection are published in different kinds of conference proceedings and journals. The conferences cover approximately 21% of the research articles, journals cover 73% and the remaining 6% are included in thesis and online reports as shown in Fig. 2b. The maximum percentage of research publications have been taken from journals, followed by conferences.

Fig. 2
figure 2

a Year-wise publications on food adulteration detection methods. b Status of publications from different sources

Full size image

5 Preliminaries for Food Quality Detection

5.1 Dataset

The first phase to perform food adulteration detection is the dataset collection. Mostly the researchers have created their customized datasets based upon their requirements and the parameters to be focused upon to evaluate the food quality. Some of the annotated datasets of images of fruits, vegetable datasets and other few datasets of food of other countries are available online as given in table Table 3. These datasets include pictures of different fruits, vegetables and some dishes. As the majority of the research work has been done for food recognition and classification, therefore these datasets are available for such purposes. The information given in Table 3 helps in attaining the answer to the research question RQ2. This table provides a summary of the online available annotated datasets for different food items.

Table 3 Summary of annotated datasets for food quality detection
Full size table

6 Techniques to Detect Food Adulteration

The food adulteration detection techniques vary from simple visual methods to complex systems. Quality control tests for fruits, vegetables and dairy products are considerable aspects to assure adulterant free products for consumption.

To address the research question RQ3, it has been analysed that food adulteration detection techniques can be broadly classified into two categories as shown in Fig. 3. These are Conventional methods to detect food adulteration and Automated food adulteration detection techniques. These techniques have been discussed in detail in the following subsections.

Fig. 3
figure 3

Methods to detect food adulteration

Full size image

6.1 Conventional Methods to Detect Food Adulteration

The traditional methods used to detect food adulteration include simple chemical tests, check the freshness of food based on smell, various electronic devices or some other manual and observation-based methods. Some of the techniques followed to detect food adulteration are discussed below.

  • Electronic devices to identify food adulteration: It has been observed that various digital devices that are used to identify the adulteration of the food like lactometer test that measures the specific density of the milk, freezing point determination of the milk or lactoscan device which is used to perform digital analysis of the milk etc.

  • Laboratory tests using chemical substances: The various laboratory tests that are conducted to check the adulteration of the substance. For example, various pH indicators are used that change the colour of the food if an adulterant ingredient is present in the product. Various chemicals are used to find if the food item is adulterated or not. For example, tincture iodine is used to detect the presence of starch in milk and milk products [43]. Hydrochloric acid (HCl) is used to identify washing soda in jaggery. These lab tests are conducted in the presence of the experts with high domain knowledge and who know about the chemical properties of the food products.

  • Manual inspection of food products: This method involves the human intervention to manually perform checking of the food products to detect adulteration. Human experts are needed to independently check each adulterant and outline their properties. The manual inspection is done based upon the physical appearance of the food products, their smell, texture and other such parameters. In some instances, even these experts would have trouble making appropriate predictions.

  • Observation-based methods to detect common food adulteration: There are different methods based on observation which are used at home and other small scale levels to identify the adulterants in food products such as to detect the water in milk put a drop of milk on the polished slanting surface and if the milk flows leaving the trail behind the milk is pure and in case of adulterated milk, it flows immediately without leaving any mark. Similarly, to detect other oils in coconut oil, refrigerate the oil in transparent glass and the coconut oil solidifies and other oils are deposited as a separate layer in the glass. There are many other methods like heating and dissolving the samples in water etc. are followed which detect the adulterants in different food products.

6.1.1 Challenges of Conventional Methods of Food Adulteration

Researchers have analysed that an automated system sometimes identifies food quality better than humans do. Automated systems can play an important tool in evaluating the quality of different food products. These traditional methods are quite complex and require a lot of manual work and hence there are chances of delay as discussed above. Also, the specialized infrastructure is required for accurate analysis. Not only this but also various agencies are needed for the verification and validation of the results. These methods demand a long time and are quite tedious and inefficient. The current methods of detecting potential adulterants tend to be a burden on domain experts, who do not have the resources and capabilities. Hence to improve these methods and to increase their accuracy; AI comes into a picture that employs different techniques to detect the food adulteration.

6.2 Automated Food Adulteration Detection Techniques

AI has been playing a predominant role in the world of food safety and quality assurance. The food industry has been leveraging the advantages of the latest advancements in AI. Product inspection in the food industry is in high demand, including quality inspection of the food products, classification and grading of fruits and vegetables.

To build an AI system to detect food adulteration, the data required for analysis can be in two forms. The first type of data is used to develop a vision-based model where the main focus is to inspect the quality of food items based upon the parameters like colour, texture, size, shape, defects, morphological features etc. The classification is done based upon the physical appearance of the fruits, vegetables and dairy products by the training of the data using ML models. The second data type consists of parameters such as moisture, pH, temperature, pressure, humidity, viscosity and other related parameters which aim to examine the chemical composition of the food products. For this purpose, the Internet of Things (IoT) can be used to collect the data from different sensors and the decision of adulteration detection is done based upon the contents of the food products. There are other modern types of equipment like electronic tongue and electronic nose that acquire such data and perform analysis to evaluate the food quality.

6.2.1 Vision Based Methods

The visual aspect is one of the most significant parameters to assess the quality of food. Vision based model is a method in which images of the food products can be used to detect adulteration by training the visual dataset on the ML and deep learning (DL) models as depicted in Fig. 4. These models are also used to evaluate the food quality based upon the textual data. The computer vision system is used for automatic and precise classification of food products based on the adulteration level. Food quality evaluation can be done by acquiring the data through a digital camera and then ML models can be used for prediction purposes. The Artificial Neural Networks (ANN) has been highly popular for research works in recent years due to their ability to learn about the features with minimal prior knowledge of the data and independent feature selection. These networks are inspired by the working of the actual neural system of the human body that works like a black box and adjusts the weights as per training rule. The neural network is used in those cases in which the exact model is not known. The ingredients of the food products are used to train the neural network using images to extract textural, shape and morphological properties. The successful implementation of ANN classifiers can be done to inspect the quality of food products and for grading purposes.

Fig. 4
figure 4

Automated Methods to detect Food Adulteration [44]

Full size image

Convolution Neural Network architecture can be proposed that is capable to extract features to detect the food adulteration. The approach based upon a vision system is considered as eco- friendly as it does not involve the usage of chemical reagents. There are several other advantages of this method such as quick and efficient results, low cost and can be widely used in the food industry for classification and grading purposes. To answer the research question RQ3 the details of the vision based methods used by researchers have been described in Table 4.

Table 4 Summarized review of Food Adulteration Detection System using vision based methods
Full size table

Few researchers have used these machine learning methods to access the food quality for textual data. The dataset consists of different parameters related to the food product. Most of the researchers have used ANN to perform the detection of food adulteration. The details of the methods used in the literature survey are summarized in Table 5.

Table 5 Summarized review of Food Adulteration Detection System using machine learning
Full size table

6.2.2 IoT Based Food Adulteration Detection Methods

The AI along with the Internet of Things (IoT) has been a powerful platform for food security and safety. IoT technology can be used to make the system of adulteration detection as a smart device. To enhance the quality of food, it could be a part of the food supply chain by tracking food conditions and live sharing the data with the consumers. As shown in Fig. 5, the system works in three phases: (i) Sense (ii) Analyze (iii) Predict.

Fig. 5
figure 5

Working of IoT System

Full size image

In phase one, the data is collected using different sensors. The different sensors can be used to record the data and a Raspberry Pi can be used to control the entire working of the system as shown in step 1 and step 2 in Fig. 5. Next, in phase two the design and implementation of an AI system which may use the data collected from sensors. The food quality system can be designed to keep a watch on various environmental factors such as temperature, humidity, alcohol content and exposure to light that may decay the food as shown in step 3 and step 4 as per the figure. Finally, in phase three the smart decisions for the food adulteration detection system are suggested and predict the output if the food product is pure or adulterated as seen in step 5 and step 6 in the figure. The quality of the food can be detected which can be used as a platform for the food adulteration monitoring system. The related work done by researchers using IoT to perform food quality evaluation is summarized in Table 6.

Table 6 Summarized review of food adulteration detection system using IoT
Full size table

6.2.3 Quality Control of food using Electronic Methods

The electronic nose (E-Nose) and Electronic Tongue (E-Tongue) are devices that work the same as human nose and taste organs and are composed of an array of sensors. These systems have broad applications in the food adulteration detection system as the complex data sets from E-Nose and E-Tongue signals coupled with multivariate statistics constitute fast and effective instruments for classifying discriminating, recognizing and identifying samples, as well as predicting the concentrations level of various compounds.

An electric nose is used to detect the smell even better than the human’s sense of smell. It obtains the data on the nature of the compounds under consideration by using the chemical detection principle; it is a smart sensing tool that utilizes the array of gas sensors that overlap with the pattern of reorganization component. The detection system of the e-nose consisting of sensors when comes in contact with the volatile compounds experience a change in electrical properties. A specific response recorded by the electronic surface transforms the signals into digital values [78]. Computation is done based on the statistical models on the recorded data. It is widely used in the research fields as it detects the hazardous gas present in the adulterated food that is not possible for a human nose.

E-Tongue is a multichannel taste sensor that is used to recognize, classify and quantify the components of liquid samples. The data about the samples are collected from sensors that work the same as gustatory cells present in taste buds of the tongue. The set of specific sensors is used to obtain the digital fingerprint of the sample and the information related to taste generating substances is transmitted into electrical signals which serve as a profile input for the data recognition system. Figure 6 presents the working schema of the E-Nose and E-Tongue where it is observed that the food products can be accessed in three forms solid phase (fruits, vegetables etc.), liquid phase (milk, juices etc.) and gaseous phase (volatile compounds emitted from the products). The solid compounds are analyzed through their texture, shape, colour, thermal and optical properties whereas in case of liquids the taste of the food item is analyzed and for gases, the compounds are judged based upon their odour and volatile compounds emitted from the substances. These electronic devices mimic the olfactory system of the nose to study the gaseous elements and gustatory receptors of the tongue to look over liquid compounds.

Fig. 6
figure 6

The architecture of E-Nose and E-Tongue [79]

Full size image

The sensory devices such as a spectrophotometer, thermometer etc. are used to examine the solid products. Further, ML and DL models can be applied for classification and pattern recognition based upon which the prediction is done about the presence of adulterants in food. Also, the other parameters such as odour, taste, the flavour of the food, aroma appearance and texture of the food can be analyzed from the predictions of the model. The research work related to these electronic methods is described in Table 7.

Table 7 Summarized review of food adulteration detection system using E-Nose and E- Tongue
Full size table

Thus, to build a computer-based food adulteration detection system, one needs to use IoT devices to sense the data and ML models for predictions based on the collected data.

7 Machine Learning and Deep Learning Applications in Food

Machine learning is a subfield of artificial intelligence that makes computers learn without being explicitly programmed. It is constructing the algorithms that can learn from data and make predictions on the related data. ML is categorized into two classes namely supervised and unsupervised machine learning. In supervised ML, there is a predetermined set of classes into which the food items are classified and training data is available for each class. The system uses any of the classification algorithms such as Naive Bayes (NB), Support Vector Machines (SVM), Decision Tree (DT), k-Nearest Neighbour (k-NN) and trains a model from the given data. This trained model is then used for making predictions and assigning the food products into different categories. In the case of the unsupervised approach of ML, no labelled data is provided to models. Deep Learning is a part of machine learning that can be considered as a representation learning method that is inspired by the human brain. It has turned out to be a powerful medium of pattern recognition in the last few years. It uses a deep neural network composed of multiple layers of neurons. It has an advantage of self-feature engineering and good accuracy that makes it work efficiently for even highly complex problems. Deep Learning models are good for classification and regression tasks provided a sufficient amount of data. Deep Learning signifies the number of layers that contain the neural network. The primary neural network is composed of three layers input, hidden and output layer. The number of intermediate layers increases as the complexity of the problems arise.

The inclination towards machine vision in the food domain is quite trending from the last few years. It has experienced a huge expansion in both theory and implementation. It has various applications such as medical diagnostics, automated manufacturing, aerial surveillance, remote sensing and now grading of food and agricultural products. Object detection is one of the classical applications of computer vision that is involved in developing an automated food sorting system. In this, the objective is to identify what and where i.e. to recognize the objects in the given image and where the objects are located in the given image. The task of object detection is a bit complex as compared to object classification as it just recognizes the objects but not their location in the image and also classification fails in images having more than one object. Convolution Neural Network is also used for object detection. It is a deep learning concept in which an image is given as the input, features are extracted from the same and based upon these features, the input image is differentiated from the other images. The answer to RQ4 is discussed in the following sections that tell about the applications of machine learning and deep learning in the food domain.

7.1 Food Recognition and Classification

Eating habits and daily diets can highly affect the health of people. Food Recognition and classification is a major task that helps human being to record their daily diets. It is mandatory for diabetic and allergic patients to strictly monitor and control their dietary behaviour. The information about the food products is characterized in their images. Image acquisition is an easy and cost-friendly medium for procuring information for food recognition and classification. This task becomes challenging for natural products that vary in size, shape, volume, texture, colour and composition. Various background and layout of foodstuffs also introduce variations for food recognition and classification. At present, due to the common use of CNN, image analysis has been the most commonly used pattern in food recognition and classification.

There are various popular machine learning and deep learning methods for this purpose. But, CNN architectures work better for image identification. These structures include AlexNet [89] [90], a network using repetitive units called visual geometry group network (VGG) [90], GoogLeNet [91] that includes parallel data channels, and residual neural network (ResNet (He et al. 2016) constructed by residual blocks [91,92,93]. Furthermore, these mentioned network architectures can be downloaded from the model zoo with pre-trained weights. That is, the models have already been trained by some image datasets like ImageNet [93] so that these pretraining models have already learned the ability to extract image features (such as colours, texture information, high-level abstract representations, and so on) [94]. Researches can use their specific image datasets to implement transfer learning based on the pre-trained model, which means that we can use our dataset to retrain the weights of a fully connected structure for final classification while keeping the weights of convolution layer unchanged, or slightly adjusting the weights of the whole network. The researchers who have worked upon food recognition and classification are listed in Table 8.

Table 8 Summarized review of food recognition and classification
Full size table

7.2 Food Calorie Estimation

With improved living standards, dietary management is gaining more and more attention. Nowadays, technology can support users to keep track of their food consumption in a more user-friendly way allowing for a more comprehensive daily dietary monitoring. People are more cautious about keeping track of their daily diet to help them control nutrition intake, lose weight, manage their diabetes or food allergies, and improve dietary habits to stay healthy. The food calorie is one of the most concerned indexes. Many mobile APPs have been designed for recording everyday meals including not only food names but also food calorie [101, 102] [103, 104]. The task of food calorie estimation is more challenging than food classification because it’s not sufficient to estimate the food calorie just from its texture and colour. The weight or volume of food, cooking directions and ingredients directly affect the calorie content of the food. It is not easy to build a large dataset containing food images, ingredients, cooking methods, and weights (or volumes) labelled with calorie content, which restricts the use of deep learning technology to achieve calorie estimation. Few authors have worked on the image classification part but it just roughly estimates the calorie content of the food. The research work in this area is presented in Table 9.

Table 9 Summarized review of food calorie estimation
Full size table

7.3 Food Supply Chain

The food supply chain is a complex system consisting of multiple economic stakeholders from primary producers to consumers (including farmers, production factories, distributors, retailers, and consumers) [103] [105]. It is hard for regulators, such as governments, to obtain reliable food information due to the unreliable information from the supply chain, which can easily lead to food fraud and food safety problems. Mao et al. [103] presented a credit evaluation system based on blockchain for the food supply chain using a deep learning network named LSTM. The evaluation task was carried out by analyzation of credit evaluation texts. Text data like “The fruit does not look very fresh” were labelled as “negative,” and the sentence such as “the quality is good” has a “positive” label. Sentence feature extraction was performed by LSTM and these features act as the input of DNN based classifier. The proposed method showed approximately 90% classification accuracy on a Chinese text dataset, which was beyond the reach of traditional methods such as SVM and naive Bayes. The research in [103] solved a problem that how to transform a large number of credit evaluation text data into some simple evaluation indicators [105]. Similarly, another article [104] [106] introduced a sentence classification method using deep learning, which can be tested using the dataset mentioned in Mao et al. [103] for comparison. Such research relies heavily on big data, so much work (such as dataset in the type of audio, text, and so on, for food domain should be collected, for example) remains to be done in the future.

7.4 Quality Detection of Fruits and Vegetables

Fruits and vegetables being an essential part of a healthy diet are explored a lot and their quality detection has been a trending research area. It provides the necessary nutrients for human beings. Deep Learning and machine learning along with image processing has become an efficient tool for fruit and vegetable quality detection in the last few years. The various research studies on the quality detection of fruits and vegetables are presented in Table 10.

Table 10 Summarized review of food quality detection of fruits and vegetables
Full size table

8 Findings of Systematic Survey

This section summarizes the established findings as the comprehensive survey has been conducted and efforts have been made to answer all the research questions given in Table 1. From the literature survey, it is concluded that the researchers have used machine learning and deep learning techniques followed by sensor-based approaches as shown in Fig. 7. PCA has been commonly used for dimensionality reduction. There is too much potential in AI-based methods, beyond some of the labour intensive tasks for food adulteration detection. Most of the research experiments (i.e. approx. 75%) have therefore opted for machine learning as well as deep learning in parallel to detect food adulteration. Nevertheless, researchers are being attracted in recent times to experiment with deep learning due to improved accuracy regardless of the time constraint needed to train the data. Researchers have also used sensor based methods and primarily focused on ANN and CNN for the development of a food adulteration detection system. The response to question RQ3 has been obtained through Fig. 7 which summarizes the AI-based techniques for food adulteration detection.

Fig. 7
figure 7

Percentage of research work using food adulteration detection methods

Full size image

It has been observed that the food products that have been focused by the researchers are dairy products mainly milk, cheese and ghee. In the category of fruits, the experiments have been conducted on apples, mangos, oranges, bananas, grapes and plums etc.

The vegetables for which adulteration detection and quality evaluation have been performed are the carrot, tomato, onions, cucumber, spinach etc. Also, the researchers have worked to detect adulteration in other food products like honey, olive oil, saffron etc. Figure 8. depicts the commonly adulterated food products. The response to research question RQ5 is stated in Table 11 which describes the major food items that have been explored by researchers to study about food quality detection system.

Fig. 8
figure 8

Percentage of food products commonly used for adulteration

Full size image
Table 11 Commonly used food products for food adulteration detection
Full size table

8.1 Features used for Food Quality Detection System

The research articles have been studied thoroughly and the answer to the research question RQ6 has been reported in Table 12. As discussed before that the food item is evaluated based upon vision parameters that include colour, morphological and texture features.

Table 12 Different parameters studied for food adulteration detection
Full size table

8.1.1 Colour Features

Colour features are the ones that influence the buyer to accept/ reject the foodstuff as it characterizes the freshness and quality. Images of the food products are acquired by widely used RGB colour models based on red (R), green (G), blue (B) primitive colours. This colour model separates an object into red, green and blue planes. In an image, various RGB devices produce different RGB values for the same pixel, multiple transformations techniques are used to standardize these values. As RGB is not linear with the human visual inspection, the sensory properties of food products cannot be analyzed. HSI is proposed and developed to resolve such problems. It is a leading method for evolving colour based image processing algorithms that are easily visualized by humans. However, HSI and RGB are similar and do not respond to minute colour variations. Therefore, it is not advised to evaluate the transformation of product colour during processing. CIELAB colour space identifies all the colours to the human eye and was designed to be used as a device dependent model as to where ‘L’ is the measure of lightness, ‘a’ and ‘b’ adjusts the red/green and green/blue balance respectively. It can be perceptually related in such a way that colour differences in CIE-LAB space that a person recognizes are the same as Euclidean distances. Since the colour measured by computer vision can be compared easily with colour obtained from CIELAB colour space, it provides a feasible way to assess the quality of object colour measurement. Table 13. summarizes the quality analysis of fruits and vegetables based upon different parameters and colour space model employed by numerous researchers.

Table 13 Comparison of different colour features for quality analysis of fruits and vegetables
Full size table

8.1.2 Morphological Features

The morphological features that depict the shape and size are generally used to classify fruits and vegetables. Grading of these food items is performed based upon their size. The size of the food item is measured using features like projected area, perimeter, length, width, major and minor axis. Such features are commonly used in industries for automated sorting purposes. The area (a scalar quantity) measures the actual number of pixels in the region. Pixels of the area are used to calculate the projected area. Feature extraction is done using the gap of two neighbouring pixels. Perimeter is determined by the distance between the boundaries of the region. These features are efficient enough once the object is segmented, irrespective of shape and size. Length and width are used to measure the size of fruits and vegetables. As the shape of food products typically changes during processing, it is necessary to restore the orientation at which length and width are measured. The major axis is the longest line across the object, obtained by the length of each of two boundary pixels.

The shortest line perpendicular to the major axis is the minor axis. The shape is a crucial visual attribute for image content description that cannot be defined precisely because it is not an easy task to measure the similarity between shapes.

The two shape descriptor categories are region-based (based on integral object area) and contour-based (boundary segmented by local features). Roundness, aspect ratio and compactness are used to characterize the shape of the food item. In the food industry for quality analysis convexity, roundness, compactness, length, width, elongation, boundary encoding, length/width ratio, Fourier descriptor and invariant moments are the most common shape features used. Table 14 illustrates the quality analysis of fruits and vegetables based upon different morphological features employed by different researchers.

Table 14 Comparison of different morphological features for quality analysis of fruits and vegetables
Full size table

8.1.3 Texture Features

The texture is suitable for a wide range of objects that understand and interpret human visual systems. Texture measured from a group of pixels depicts the distribution of elements and appearance of the surface and is useful in machine vision which predicts the surface in form of roughness, contrast, entropy, orientation, etc.

The texture is consistent with maturity and sugar content (internal quality of fruits and vegetables). It is also used by extracting intensity values between pixels to isolate different patterns in images. Quantitative and qualitative analysis can be used to study the texture. As per the quantitative analysis, six textural characteristics i.e. contrast, coarseness, line-likeness, directionality, roughness and regularity whereas four features i.e. contrast, correlation, entropy and energy are according to qualitative analysis. Statistical texture, model-based texture, structural texture and transform based texture are various types of texture characteristics. Statistical texture, extract matrix which is dependent on intensity values of pixels. The different model-based texture is a fractal model, random field model and autoregressive model. Structure texture comprises of lines, edges that are constructed by pixels intensity. Spatial domain images can be derived from transform based texture. The statistical texture is widely used due to low computational cost and high accuracy. Table 15 illustrates the quality analysis of fruits and vegetables based upon different morphological features employed by different researchers.

Table 15 Comparison of different texture features for quality analysis of fruits and vegetables
Full size table

Features based upon chemical composition can also be used to detect adulteration in the food as they aim to know about the content of the food product. The quality and adulteration content of the food item can be predicted by using sensors that measure temperature, moisture, pH, volatile gases like carbon dioxide, oxygen, odour, pressure, viscosity etc. There is a particular range of temperature which is to be maintained to increase the shelflife of the foodstuff. The acidic level of the item can be measured using a pH sensor and also the concentrations of these gases change for the fresh and the spoiled food. The values of these parameters vary once the food is spoiled or the quality has been degraded.

8.2 Work in Progress

The research in this field of food quality is going on worldwide. The answer to RQ.7 is addressed below that discusses some existing studies.

  • pH Sensing Using Electrospun Halochromic Nanofibres: The researchers at IIT Hyderabad have been working on a project to develop a smartphone-based system that is equipped with sensors to detect the amount of adulteration in milk [58]. Initially, they have developed a system to measure the acidity of milk through an indicator paper that changes colour based on the level of adulteration. Besides this, they have also developed algorithms that can be incorporated on a mobile phone to detect the colour change accurately.

  • Paper Strip-Based Tests: An innovative kit has been developed to detect the adulteration of milk by the National Dairy Research Institute (NDRI), Karnal [165].The paper strip-based tests have been developed which can rapidly detect adulteration of milk containing neutralizers, urea, glucose, hydrogen peroxide, sucrose and maltodextrin. The test involves dipping a strip in the milk sample for a short duration followed by immediate visualization of the colour of the strip.

  • Food Sniffer: A smart portable kitchen gadget has been developed to check the freshness of raw meat, poultry or fish. It’s a wireless device designed by Swiss Scientists that detects these non-veg products if they are fresh, spoiled or in the stage of getting spoiled and the results are displayed on the smartphone [167]. It contains a sensor that collects the gases emitted by meat to examine its freshness and helps to avoid food wastage and of course takes care of food safety.

  • Paper Sensor to detect Milk Freshness: The researchers at IIT Guwahati have developed a paper kit that can test the freshness of milk and tell how well the milk has been pasteurized [168].They have used a filter paper to design a detector that can react with a milk enzyme, Alkaline Phosphatase (ALP). This compound is removed from milk once it is pasteurized. Hence, ALP acts as an indicator to milk quality because it reacts with the sensor probe to generate precipitate and indicates the presence of microbes in the milk. The colour of the paper on dipping in milk changes and these colour changes are captured using a smartphone camera to get the corresponding RGB values. These values are compared with the standard threshold values to measure the amount of ALP present in the milk sample and the quality of the milk can be predicted based upon the analysis.

  • Fruit Sorting at Amazon: There has been a practical application of machine learning at Amazon that uses these algorithms to predict the quality of groceries. It grades different types of products and prevents the wastage of fruits and vegetables by providing consistent results. It predicts if the fruit quality is good or bad. The different fruits stored in the warehouse are scanned through a set of cameras and sensors to inspect their quality.

The research question RQ8 is replied in Sect. 7. The key results of the research questions mentioned in Table 1 from this systematic survey can be summarized as follows.

  • The research work in the field of food adulteration detection is being done for a long time but we have studied the research work of the last decade to gather information about the current trends followed to evaluate the food quality.

  • The annotated datasets are available for different food categories like fruits, Italian and Japanese food items. Researchers can easily use these resources as a description along with online availability is provided in this systematic review.

  • Different food adulteration detection methods such as machine learning, deep learning and IoT based detection systems are briefly described in this research study. However, due to better accuracy achieved by these techniques, the researchers are also enticed by deep learning techniques

  • This survey brings out major machine learning and deep learning applications in the food domain.

  • This systematic review comes up with the information about the food items that have been explored by researchers to work upon to detect the adulteration and to evaluate their quality.

  • From this systematic survey, it can be said that researchers have tried to work upon colour, texture, defects, and morphological features etc. Also, they have focused upon features that try to extract the components of the food products.

  • It has been noticed that there are few existing similar works going on who are working to develop a food adulteration detection system.

  • This research survey concludes that researchers have performed mostly on food adulteration detection for the presence or absence of adulterant.

It has also been seen that an extensive amount of research work has been carried out for food adulteration detection using various approaches and techniques. Therefore, these studies can also be implemented to develop a low cost, user-friendly food adulteration detection system. To perform the food adulteration detection system for the Indian market, one can generate his dataset and can develop a model using the available features and models. Also, the best performing ML technique used by researchers for food adulteration detection can be applied to the Indian context dataset.

9 Conclusion and Future Work

We were inspired to conduct this systematic survey by the growth of research work in the area of food adulteration detection. The extensive research has been done to keep a quality control check on food but still, the Indian market is facing major challenges due to less technical development in the field of food adulteration detection. Certain challenges are witnessed while building an AI and Sensor-based system to detect food adulteration.

Few research studies are available that cover a thorough analysis of the machine learning techniques to detect food adulteration. We realized the need for a systematic literature survey after reviewing the groundbreaking research in the field of detection of food adulteration. This paper is, therefore, an important contribution to the literature on the detection of food adulteration using artificial intelligence. This survey includes 112 research studies published on food adulteration detection using machine learning from the last decade to include the relevant work only. The 112 research studies taken into account in this systematic survey have been determined by developing a review protocol that includes the research questions, sources of information, inclusion and exclusion criteria. The different results of this survey have been examined to get the answers to the targeted research questions drafted in this article.

The overview of the food adulteration detection approaches, major food products to be worked upon, and important parameters are given in this paper. From this review, it has been observed that most of the research work in this field has been published in conferences followed by journals. It has also been analyzed that mainly ML and DL (i.e. 75%) approaches have been used by the researchers have in comparison to other sensor based and hybrid approaches. This paper also gives about the different types of experiments followed by researchers to detect food adulteration. The different food categories and major parameters that can be referred to are mentioned in the paper.

There is a lot of work to be done to improve the accuracy of the detection system for food adulteration. It has been observed that the datasets are not readily available online because the researchers do not provide any links to the same. As there is a lack of annotated datasets and the creation of labelled datasets for different foodstuffs is a time-consuming task. In the future, these resources can be provided to utilize these by other researchers so that they can focus only on enhancing the efficiency of the system by developing new food adulteration detection methods. It has also been analyzed that deep learning approaches for food adulteration detection are in demand. Hence, researchers can experiment with these approaches to achieve improved results. And, also there is a need to build online systems which can perform food adulteration detection. The research can be done to develop a low-cost smartphone-system to detect food adulteration which can serve as an aid to end-users for their quality satisfaction. This field needs to be integrated with a real-time system for food adulteration detection. Thus, this survey can help the researchers in building the effective food adulteration detection system by using the different methods and techniques used by other researchers which can help in the benefits of society.