Keywords

1 Introduction

In this present time, people are more considerate about their health and diseases and it compels them to be more conscious about the everyday food and the diet. Not only about the good food and diet but people are considerate about the nutritional values that are contained within a food. Technology has touched almost every aspect of human lives with its efficient applications and techniques and with the exponential growth of technology and evolution of technology, the traditional method to classify food has been replaced with applications that automatically detect the food and recognize their nutritional details from the pictures captured using different machine learning algorithms and computer vision. Applications can automatically scan the diets of individuals and help in numerous aspects.

Overeating is concerning these days because people are overeating and it makes them less active. Considering the busy schedules and stressed lives of people, the importance of proper classification of food is vital and will play a significant role in the lives of people.

Over the past few years, a fair amount of research and development have been carried out in the field of calorie analysis and visual-based diet, and still, the efficient and structured extraction of information from the food clicks remains an exigent issue. Few of the techniques that are currently in use for dietary assessment included manually recording instruments and self-reporting and doing it manually makes it a tedious task to execute. To overcome this, enhancements to the present techniques are a necessity. One of the possible potential solutions to overcome this challenge is the mobile cloud computing system.

In this paper, the YOLO algorithm has been used to classify the food images. The main praise worthy feature of this algorithm is its remarkable speed. It is outstanding when it comes to execution and speed and processes 45 frames per second. With the capability to acknowledge generalized object representation, YOLO algorithms stand to be the best algorithm for detection of objects. The architecture of this algorithm is more like fully convolutional neural network (FCNN). Full images are trained by YOLO, and it precisely optimizes the detection performance.

1.1 Organization of the Paper

Section 2 that is literature survey presents the related works and techniques that are used in classifying the food through the processing of images. Section 3 details the proposed methodology using the YOLO algorithm, and further in the paper Sect. 4, results and analysis have been explained. Conclusion and future scope have been presented in the Sect. 5 of the paper, and Sect. 5 marks the end of the paper.

1.2 Contribution of the Paper

  • In dataset used in this work consists of 80 different categories or classes of Indian Food Images consisting of 4000 instances.

  • The paper explains the YOLO algorithm in detail along with its use in developing the food classification model.

  • Comparison based on efficiency and working methodology has been done among several algorithms and techniques that are used in developing food classification and presented.

  • Recent work of many different researchers focusing on food classification has been explained in the literature survey of the paper.

  • The paper also provides an insight on the future work that could be done in order to enhance the performance of the model that is presented in the paper.

2 Literature Survey

In this literature survey, multiple papers targeting food classification using image processing techniques and many different algorithms have been reviewed, and the information extracted after reviewing the papers has been mentioned below.

To develop the model for classifying food using the food images, in [1], the dataset that was used contained 101,000 images and 101 categories. To make the system realistic, this dataset was considered. In the dataset, each food category contained 750 clips for training and 250 clips of testing. To train the huge dataset that contains multimedia data, CNN requires high-performance computing machines. After training the system properly, it was able to produce results in an efficient time.

The model proposed in [2] is divided into three contrasting parts. The first part is pre-trained convolutional neural network nodel, the second is dataset preparing and pre-processing phase, and the third and last part is textual data model training. Information’s like the type of the food and its attributes such as nutritional value and caloric value is provided by the system proposed in [2]. Image of the food is taken by the system, and the image then is classified. After the classification of the image, the system details the attributes of the food. Further, the result is enhanced utilizing multi-crop, data augmentation, and similar technologies like these. The model proposed in [2] achieved the exactness quite well, and an accuracy rate of 85% was achieved.

According to [3], the dataset that was used for building their system was the publicly available Food 101 dataset which has 100 images of 101 classes. Further, for the classification of these images, SVM was used. Average accuracy was reported after performing fourfold cross validation. In the system that is proposed in [4], although the dataset consisted of 101 classes but only 50 classes were used in the actual work. To store the missing information, BDF and GPCA were used. To extract the feature, LBP and NRLBP were used. They were fed into SVM classifiers for identifying food images. The accuracy obtained for the proposed model was not mentioned in [5].

In the paper [6], personalized classifiers are expanded on a large scale for daily food image identification in the real world. The architecture of the model comprises a NCM classifier, and the other classifier which is used in the architecture is NN for each user and a model of food distribution which is time independent has been used in order to achieve better performance and exactness in the result.

According to [7], the model has used convolution neural networks to train the dataset, and at the end, the accuracy of 61.4% and top accuracy of 85.2% have been achieved. The dataset used was Food 101 dataset and was trained from scratch. ImageNet weights were used to pre-train the models. The model that outperformed all the other models was pre-trained InceptionV3 model whose top layers were unfrozen in stages.

3 Proposed Methodology

The algorithm used to train the model is the YOLO algorithm. Image processing using the YOLO algorithm is considered uncomplicated and straightforward. The You Only Use Look Once (YOLO) algorithm is capable of training on full images, and it directly optimizes the detection performance. It has numerous benefits over the regular traditional methods. The design of YOLO algorithm permits end-to-end training and real-time speeds and maintains high-average precision. The YOLO algorithm is based on regression, it does not select any particular part of the image it rather predicts the bounding boxes and classes for the full image in a single run of the algorithm.

Instead of searching for the interested regions in the image which is being inputted and could contain an object, YOLO algorithm splits the input image into numerous cells and each cell becomes responsible for prediction of K bounding boxes. YOLO signifies the probability that the cell holds a particular class. The equation for the very same is

$${\text{SCORE}}\,m,n = Pm*Mn$$

Probability of presence of an object of certain class ‘m’.

YOLO is said to be a clever and convolutional neural network (CNN) and is known for doing object detection in real-time. Single neural networks are applied by YOLO algorithm to the full image, and then, the image is divided into regions and it predicts probabilities for each region and bounding boxes. Predicted probabilities weigh the bounding boxes. The General Yolo-based detection system is depicted in Fig. 1.

Fig. 1
An image depicts the general Yolo-based detection system.

YOLO detection system [4]

Full size image

The dataset that is used in the following model is a self-prepared database, and it consists of 4000 different images and 80 different types of food as depicted in Fig. 2. Sample images used for training the model are depicted in Fig. 3.

Fig. 2
Snapshot of the 80 classes available in the dataset.

Names of 80 different classes available in the dataset

Full size image
Fig. 3
Photographs of the food sample images in the dataset.

Sample images from the dataset [8]

Full size image

The model that is proposed for food classification in this work is depicted in Fig. 4 and elaborated as follows:

Fig. 4
An image depicts the working flow of the proposed model. 1. Image collection. 2. Training with the Yolo algorithm. 3. Image classification.

Steps involved in developing the model

Full size image

  • 4000 images of forty different types of food were captured.

  • After using the dataset that contained 4 k different images of food, image encoding was applied for all 80 classes for which LabelImg tool was used.

  • Dataset was divided in 70–30 ratio for training and testing the model.

  • Finally, to train the model, YOLO algorithm was used with 6000 epochs.

4 Result and Analysis

The proposed system in the paper yields a noticeable accuracy rate of 99%. The YOLO algorithm which is used to train the model uses a totally different approach. The extremely fast speed of the algorithm makes it more popular, and the additional benefit that comes along with this algorithm is its capability to run in real-time.

The proposed model has innumerable benefits over other methods and that is as follows:

  • The YOLO algorithm is extremely fast.

  • It looks uses the encoding of the image for training and testing.

  • It is comparatively easier in implementation.

  • It outperforms various other detection methods.

The above-mentioned attributes of the YOLO algorithm used in training the model helped in achieving a decent and remarkable accuracy rate. The results obtained after using various techniques which were used for training the models are given in Table 1. Algorithms like ResNet-50, VGG-16, ImageNet, Inception, and YOLO were used for training the model. The dataset was divided into 70–30 ratio for training and testing purpose, respectively, as depicted in Fig. 5. For generating efficient training and testing results of the proposed model high-configuration architecture consisting of AMD RYZEN 9 4000 Series processor, 64-bit Windows 10 Operating System, and 32 GB of RAM, NVIDIA GeForce GTX 960 was used.

Table 1 Comparative analysis of the results obtained using various techniques
Full size table
Fig. 5
A pie chart depicts the dataset specifications. 1. Training data 70 percent. 2. Testing data 30 percent.

Dataset specifications after splitting the data

Full size image

5 Conclusion and Future Scope

With the emerging need of classification of food based on their nutrition and various other parameters, traditional methods prove to be extremely inefficient and a time taking process. With the evolution of technology over a long period of time, researchers have found various methods to classify food in a more efficient way. In this paper, the YOLO algorithm has been explained in detail, and the paper also details the methodology to build a system that classifies food using image processing techniques which uses the YOLO algorithm to train the models. The system proposed in this paper gives a remarkable accuracy rate of 99%. Numerous papers that focus on the image processing technologies have been reviewed and elaborated in the literature survey of the paper. We can achieve more accuracy if the dataset is precise and contains more unique images of types of food. Impurity of the training images needs to be removed for getting more enhanced performance. Also, the number of epochs used for training the model can be improvised for getting more promising results.