1 Introduction

Underwater imaging plays an effective role in ocean exploration but often suffer from severe quality degradation due to artifacts and light absorption. Today, most of the underwater vehicles used for underwater exploration are usually equipped with optical cameras to capture visual data of underwater objects, shipwrecks, coral reefs, pipelines and telecommunications lines in the seas and oceans, etc., [1, 2]. For example, for regular inspections of oil and gas pipes in the seas and oceans, ROVs equipped with these cameras are recruited with the human operators onshore analyzing images transmitted by these devices [3,4,5]. However, the color images captured with these cameras, due to the physical properties of the aquatic environment, have non-real-time manner, poor visual quality, opacity or luminosity with a poor field of view. The light is exponentially attenuated when travelling through the water, causing these images to have low contrast and opacity [6,7,8,9,10,11,12]. Absorption, which diminishes the energy of light, and scattering, which alters the path of light radiation, are two main factors responsible for underwater light attenuation [11, 13]. These floating particles, known as “sea snow” [14], are of diverse type and concentration, increasing the degree of absorption and scattering. Another challenge in this area is the expanded range of vision by artificial light that concentrates light at the center of the image rather than its edges [8]. Another issue is the greenish or bluish color of images taken underwater.

The real-time image processing techniques is always in high demand for multiple requisitions used in automated systems such as remote sensing, manufacturing process and multimedia applications; those required to have high performance. Based on that requirement, image processing and machine learning systems have been proposed in some studies using real-time concepts for underwater imaging [68, 69].

Underwater light attenuation restricts the vision range to about 2 m in clearwater [15, 16] and to less than 2 m in murky water [17, 18]. As we go deeper into the water, the light is dimmed, and the colors disappear in the order of their wavelength [19,20,21]. For example, red disappears at a depth of 3 m, and as we plunge into the water, orange starts to fades until it completely disappears at a depth of 5 m. Then, yellow mostly disappears at a depth of 10 m and eventually green and purple vanish at more profound depths. In addition to the above, low resolution of images will probably complicate the detection process.

Many methods developed to overcome the challenge of the low resolution of underwater images. Nevertheless, some inherent limitations such as artifacts, blurring, and noise, have highlighted the importance of developing methods and strategies for underwater photography and consequently image processing. One of the methods that has recently received growing scholarly attention for optimization of image quality is a super-resolution method that can yield promising results.

For the rest of this research paper, the organization is as follows. In Sect. 2, we review some related works including real-time techniques for underwater applications. Section 3 introduces the proposed method and describes its related aspects. Section 4 presents the results of the simulation in the form of quantitative outputs and qualitative features. Practical results will be discussed in Sect. 4.5 of the research and finally conclusions are given in Sect. 5.

2 Related work

Noise is a major cause of poor image resolution. Noise reduction is the process of removing noise from the main signal. In the underwater images, in addition to artifacts such as noise, there are other parameters such no inadequate light, noise and low contrast, non-uniform lighting, shades, suspended particles and other disturbing factors [22, 23].

In images [24,25,26,27], light wavelength compensation -physical and non-physical models [28, 29] and polarization (hardware) [30, 31] have been shown.

Bianco and Neumann [29] used a non-uniform fast method to improve underwater images, which is based on the gray-world assumption in color space transformation. The color correction is commensurate with the local changes in luminance and chrominance underlined by the summed-area table strategy. Galdran et al. [25] used red channel as one of the color decomposition strategies.

In [26,27,28,29,30], the effective color correction strategy based on linear segmental transform was introduced to respond to color distortion. As an innovative contrast enhancement method, it was able to reduce artifacts to achieve low contrast. Since most of the computations are performed for the pixel set, their proposed method was suitable for implementation and real-time execution.

In the past years, convolutional neural networks (CNN) have been the subject of increasing attention [31]. In the same vein, deep convolutional neural networks have recently been used as an inseparable part of machine learning not only in the categorization but also in other fields [32, 33]. There are, however, few studies on deep learning techniques for image retrieval and resolution enhancement. Conventional methods in the field include multilayer perceptron neural networks, which are finally interconnected and used for natural images such as the removal of noise [34] and fogging [35]. One of the methods of image restoration and image resolution improvement, which is less concerned with noise removal, is the one proposed by Cue et al. [36], which utilized auto-encoder networks for resolution enhancement processes including pipe-lines methods [37].

Super-resolution can provide deeper insights into the analysis of images. Different methods have been proposed for this purpose, each with their own advantages and disadvantages [38,39,40]. In general, there are two basic methods for correcting low resolution in images: reconstruction-based methods and learning-based methods. In the former, input images are analyzed based on the data contained in the image itself, and based on rules such as edge extension, curvature preservation and interpolation of pixels [41,42,43,44,45,46].

2.1 Real-time underwater image enhancement

Underwater absorption and scattering processes influence the overall performance of marine systems. Forward scattering (low-angle refraction) refers to the random change of light travelling from the object to the camera, which results in blurry images. Backward scattering describes the light reflected by waterborne particles on the camera before reaching the object. Backward scattering typically limits image contrast. The adsorption and scattering effect is not solely induced by the water, and other factors such as insoluble natural substances and small visible particles floating in the water are also involved in this process [13].

The real-time design of algorithms and hardware sections is necessary for the enhancement process of noisy images. Some enhancement algorithms commonly have been developed for real-time manner where include methods such as CLAHE [8], Retinex [58, 59], color correction [60,61,62,63], and classification models [64,65,66,67].

In a study [68], the Fourier transform structure is employed for images with high complexity to real-time detect targets in images. Sometimes unsupervised methods, as a separate network, not only enhance image quality but also are able to effectively real-time detect the objects appearing in the images [69]. The attempt made in this field is an unsupervised generative network to produce high-quality underwater images and to correct the color of the images.

2.2 Main focus and contribution on real-time processing

One of the main causes of image quality loss is the absorption of light by water. To compensate for light absorption; underwater cameras illuminate the ambience by means of artificial lights for image acquisition, while non-uniform artificial lighting produces shadows in the scene.

However, an alternative strategy in these approaches is an improvement in lighting and sensing systems for illumination an object before capturing the image. The time performance and computational complexity of some method will increase significantly based on these reasons. The loss of underwater image quality can be due to the absorption and backscattering of light beams. Actually, the degradation in image occurs by suspended underwater particles and the hazing, fogging, and other destructions arise from inappropriate imaging. The enhancement considers an important preprocessing operation for further processing of the image whereby, some challenges such as high dimension, good resolution, and high density of unwanted noises in underwater images increase the time of processing and computational complexity. More of traditional enhancement procedures have successfully been shown to improve the quality of underwater images, but computational complexity is high and, therefore, a limited number of methods have real-time or near real-time properties to process and enhance different underwater noisy images.

Due to the fact that the number of investigations in the field of image processing, is less than other subjects but the efficiency and use of the processing based method is high and important. One of the important manner in processing is real-time procedure that can be used to improve performance of underwater image enhancements.

The main aim of this study is to improve image quality parameters and measure, comparing the extent of image enhancement by a number of different criteria. In the proposed method, to achieve the goals, underwater images are divided into three categories: greenish images, bluish images and images of desirable color balance. Then, by introducing some changes to the new dark component method, the fogging is removed. Subsequently, we applied histogram equalization to both green and blue components for greenish images, and histogram stretching to both green and blue components for bluish images to attain overall color balance in these two components. Finally, deep learning technique is employed to improve resolution.

3 The proposed method

The proposed method consists of two phases. In the first phase, we correct the color and remove artifacts such as haze and fogging to improve the contrast-base quality, and in the second phase we use super-resolution with deep learning to improve the resolution of the underwater images. Figure 1 depicts the proposed steps to improve underwater images.

Fig. 1
figure 1

The proposed steps of underwater image enhancement

Full size image

3.1 Color correction and artifacts removing

The dark component can be expressed as (1):

$$I^{dark} (x) = \min_{c \in (r,g,b)} (\min_{y \in \Omega (x)} (J^{c} (y)))$$
(1)

where Jc is a color component of J and Ω (x) is a local space centered by x. According to the previous dark component, except for the open space (sky), the intensity of the dark component is low and near zero. The intensity of the dark component in a hazy or fogging image is a non-uniform approximation of the fog thickness, which is crucial for eliminating fog or haze; however, for underwater images, the previous dark component may be ineffective in many cases. Simply put, the dark component can be expressed as (2):

$$J^{dark(c)} (x) = \min_{c \in \Omega (x)} (I^{c} (y))$$
(2)

And the dark component I can be re-written as (3):

$$I^{dark} (x) = \min_{c \in (r,g,b)} (I^{dark(c)} (x))$$
(3)

As we go deeper, the light is diminished and colors disappear in the order of their wavelength. Among the three colored components of light, red is more significantly attenuated due to its long wavelength compared to other two components (green and blue) and in many underwater images, the values of the dark component or Idark (x) are small and close to zero. As a result, the dark component of the underwater image approaches zero. Thus, the dark component of this underwater image cannot provide information about the depth of the landscape on which a new dark component suitable for the underwater images is defined as (4):

$$J_{new}^{dark} (x) = \min_{c \in (r,g,b)} (\min_{y \in \Omega (x)} (J_{old}^{c} (y)))$$
(4)

where is called underwater dark component Jnew. Like the previous dark component, the intensity of the underwater dark component must also be low and near zero. Empirically, it can be stated that the underwater landscape backgrounds are blue in seas and oceans and green in lakes. Due to color displacement caused by the background, the intensity of green and blue components of the underwater images must be high in areas distant from the camera. Hence, the dark underwater component can qualitatively reflect the underwater distance between the viewing point and the camera. A general step involves removing hazy or fogging effect by estimating the backlight, scattering ratio, and transmission rate estimation [11].

After removing haze and fog from the underwater images, considering that green and blue images depend on color enhancement and color intensity balance, the two methods of histogram equalization and histogram stretching are used. In green and blue images, to prevent the attenuation of the red component, rather than applying the histogram equalization and histogram stretching to the three color components, it is only applied to the green and blue components and the red component is multiplied by a coefficient. This coefficient can be expressed for green images as (5):

$$\alpha_{g} = 1 + (M_{g} - M_{r} )M_{b}^{ - 1}$$
(5)

In this respect, αg is the red component coefficient of green images, and Mg and Mb are the mean green and blue components of the image after applying histogram equalization, respectively, and Mr is the mean red component of the image. To balance the mean values of color components of the image, we need to balance the greenish and bluish components as (6) [47]:

$$I_{ij} = \left\{ {\begin{array}{*{20}c} {(I_{ij} - \min_{I} )(\max_{I} - \min_{I} )^{ - 1} } & {if\,\,0 < I_{ij} < 1} \\ 0 & {if\,\,I_{ij} < 0\,} \\ 1 & {if\,\,I_{ij} > 1} \\ \end{array} } \right.$$
(6)

Subsequently, the red component is multiplied by a coefficient. This coefficient can be calculated for greenish images as (7):

$$\alpha_{g} = 1 + (M_{b} - M_{r} )M_{g}^{ - 1}$$
(7)

In this relation, αb is the red component coefficient of green images. Mg and Mb are the mean of green and blue components of the image after applying equalization histogram, and Mr is the mean of the red component of the image.

3.2 Image super-resolution

Scattered automatic coding is initially used to learn the features of images. In the second step, images from a reduced dataset can be categorized using the layers of the convolutional and pooling layer. The random correlation method is used in the proposed convolutional neural network model to facilitate the analysis of the network. The elements are selected based on their probability values in the feature map. In other words, the elements with high probabilities are selected. Unlike the maximum pooling in which the majority of elements are selected, in the mean pooling, only average elements are selected. In Fig. 2, the overall configuration of the deep convolutional neural network for resolution improvement is shown.

Fig. 2
figure 2

Deep convolutional neural network structure in super-resolution

Full size image

3.2.1 Image restoration

Conventional methods used to remove noise and blurring based on the image convolution model first construct the feature map and then applies the deconvolution to the feature map to enhance image quality with artifacts. The optimum effect that can be gained from image restoration against image twisting is that it is not necessary to determine the type of artifacts that could be applied to the image. It can also eliminate pre-processing phase before the construction of the convolution-deconvolution network. For this purpose, we use a quasi-inverse kernel to create a deconvolution process that is consistent with the kernel separation theory. The inverse convolution can be expressed as (8):

$$\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{K} * y = \sum\limits_{j} {s_{j} .u_{j} * } \,(v_{j}^{T} * y)$$
(8)

where \(\widehat{K}\) is inverse kernel and uj and vj are the jth columns of U and V, respectively. The kernel itself can also be expressed as USVT, which is actually the decomposition of single values, and sj is jth single value. Assuming smaller value for sj will provide suitable conditions for separable filters and thus \(\widehat{K}\) can be approximated for the desired settings. Hence, the main idea raised in this research is implementing the deconvolution process based on the separable kernel kernels with the aim of recovering the feature map outputs. The network can be generally assumed as (9):

$$h_{3} = W_{3} * h_{2} ;\,\,h_{1} = \sigma (W_{l} * h_{l - 1} + b_{l - 1} ),\,l \in \{ 1,2\} ;\,\,h_{0} = \hat{y}$$
(9)

where Wl is the weighted mapping of (l-1)th layer on the first lth and bl-1 is the vector bias. It is also worth noting that σ (.) is a nonlinear function, which can be either a sigmoidal function or a hyperbolic tangent. Unlike similar approaches, we used two hidden layers where the first hidden layer (h1) is created by applying kernels to dimensions smaller than the original image.

3.2.2 Loss function

To network training is conducted using a recursive descending gradient algorithm, which comprises two steps of back propagation feed-forward neural network of errors (BPFNN). The squared error function is also expressed as (10):

$$LossFunction = {{(1} \mathord{\left/ {\vphantom {{(1} 2}} \right. \kern-\nulldelimiterspace} 2})\sum\limits_{n = 1}^{N} {\sum\limits_{k = 1}^{c} {(T_{k}^{n} - Y_{k}^{n} )^{2} } }$$
(10)

where TkN and YkN are kth dimension of nth pattern of the corresponding label and the predicted image of higher resolution provided by the convolutional neural network. It should be noted that the measures of transparency as well as the normalized correlation have been used in the cost function to improve weights.

3.3 Real-time design

The approximate time of the calculations for N × N underwater images is equal to about 2n−9 + 27. In this approximation, the processed image is divided into N × N ×  64–2 blocks with a size of 64 × 64. If the image is a 16-bit 512 × 512, the processing clock cycles will be estimated as 64 Mega cycles (Mc) similar to the previous approximation. In a similar way if underwater image is an 8-bit 512 × 512, then this value is decreased to its lower limit of 32 Mc. Based on this computation, if we suppose that the computer is working to enhance the underwater images, the processor will operate to work at least at the rate of 2 GHz. Accordingly, this volume of computation is too high cost for an underwater image with today's technology, and, therefore, the image size needs to be reduced to work as a real-time process. We notice that some of the researchers proposed the upgrading operators in hardware implementation methodologies such as FPGA, DSP, and parallel architectures for different images [56, 57, 70, 71].

4 Experimental results

4.1 Datasets

Two sets of underwater photos were used to evaluate the performance of the hybrid algorithm. The first image dataset was taken from the National Institute of Ocean. Underwater images were taken from a depth of 1019 m during deep-sea mineral exploration. The ROSUB 6000 is capable of capturing images of rare underwater corals, fish, various worm species and other objects [48, 49]. The second dataset of images contained a series of images produced by PBRT, which were derived artificially from underwater images. This set is composed of numerous images, most of which are accessible [50] and used by researchers in many studies [51].

All images have three colorful channels including blue, red and green in JPG format. The dimensions of all the images were the same and the resolution of all of them was equal.

4.2 Evaluation metrics

In the first evaluation step, mean square error (MSE), peak signal to noise ratio (PSNR) and structural similarity (SSIM) measures are evaluated. In fact, as suggested by a review of previous studies, mean square error is one of the most popular and common criteria. Each of these relations is represented as (11)–(13).

$$MSE = { }\frac{1}{x.y}{ }\mathop \sum \limits_{i = 1}^{x} \mathop \sum \limits_{j = 1}^{y} \left( {p\left( {i.j} \right) - q\left( {i.j} \right)} \right)^{2}$$
(11)
$$PSNR = 10\log 10\left( {\frac{{255^{2} }}{MSE}} \right)$$
(12)
$$SSIM\left( {p.q} \right) = { }\left( {\frac{{\left( {2\mu_{p} \mu_{q} + C_{1} } \right)\left( {2\sigma_{p} + C_{2} } \right)}}{{\left( {\mu_{p}^{2} + \mu_{q}^{2} + C_{1} } \right)\left( {\sigma_{p}^{2} + \sigma_{q}^{2} + C_{2} } \right)}}} \right)$$
(13)

where x and y are dimensions of the improved images (q) and the reference image (p). In the third equation, μp, μq, σp and σq are the mean brightness of the reference images, mean brightness of the enhanced image, standard deviation of the reference image, and the standard deviation of the enhanced image, respectively. In addition, C1 and C2 are arbitrary and constant values. This evaluation index determines the position of points in the reference image that corresponds to the optimized image points. Moreover, time is another important measure considered in image optimization. To calculate the time required for enhancement of sample images with different dimensions, MATLAB software's tic-toc function was used, which offers the user an estimate of the processing time.

4.3 Results

Tables 1 and 2 show the results of the simulating underwater image quality enhancement with the aim of color correction for the first and second datasets. There are reference images in each category and the darkening, hazy, and fogging effects are applied step by step. Each of the MSE, PSNR, SSIM and time measures have been estimated in these tables. A darkening, haze, and fogging intensity of 5–40% were applied progressively to images. 200 images from the first dataset and 400 images from the second dataset were randomly selected and incorporated into the quantitative evaluation process. Although using artifacts undermined the desirability of measures, they still had acceptable values compared to various strategies. This reveals the satisfactory performance of the algorithm in the first part. In Tables 3 and 4, the results of the resolution improvement by the convolutional learning model are calculated. Some measuring methods were compared with the proposed method in terms of resolution enhancement, and this analogy was evaluated with respect to measures estimation. The image correction is assumed to be carried out in the first phase and then the super-resolution is applied by the proposed algorithm. At the time of the comparison, different image scales are analyzed and similar methods such as Bicubic, SRCNN [39], FSRCNN [40], A + [41] and finally RFL [42] techniques are compared. Overall, the outputs of the proposed method are greater than previous models.

Table 1 Evaluation metrics of enhancement step for dataset 1
Full size table
Table 2 Evaluation metrics of enhancement step for dataset 2
Full size table
Table 3 Evaluation metrics comparisons of super-resolution step for dataset 1
Full size table
Table 4 Evaluation metrics comparisons of super-resolution step for dataset 2
Full size table

The outputs of the proposed method in improving the quality of the underwater images are shown in Figs. 3 and 4 in two separate phases. These figures show the percentage increase of darkening, haze, and fogging. Figures 5 and 6 reveal the effect of the proposed super-resolution algorithm using deep learning on two image samples with four scales, respectively.

Fig. 3
figure 3

First part of image quality enhancement: The sample is an image from the first dataset where the first column contains images with 10, 20, 30, and 40% artifacts applied in left-to-right direction. The second column includes the first step of color correction and quality improvement for the first column images

Full size image
Fig. 4
figure 4

The first part of the image quality enhancement contains a sample image from the second dataset, where the first column includes images with 10, 20, 30, and 40% artifacts in the left-to-right direction. The second column outlines the first step of color correction and quality enhancement for the first column images

Full size image
Fig. 5
figure 5

Applying the super-resolution method using deep learning method: A sample image from the first dataset in the left-to-right direction for 2, 4, and 8 output scales are displayed

Full size image
Fig. 6
figure 6

Applying the super-resolution method using deep learning method: A sample image from the second dataset in the left-to-right direction for 2, 4, and 8 output scales are displayed

Full size image

4.4 Computational complexity

Data analysis has been performed in the 2019 MATLAB environment and the used system for implementation specified with Intel (R), Core (TM) and Core i7 processors with a frequency of 2.8 GHz, 4 GB RAM and 64-bit operating system. Computational complexity and time analysis in research are such that the integrated algorithm can be considered as Real-Time or near Real-Time methods. The reason for this claim is taking no much time to achieve an enhanced image with the lowest possible level of error by the proposed method. The temporal stability of the algorithm in improving image quality in both databases of underwater images is evident by applying different levels of noise. In the first set of images with dimensions of 32 × 32 by applying different noise levels, the variance level and the mean time are 2.36 s and 0.1088, respectively. The same values for images with dimensions of 256 × 256 are equivalent to 3.74 s and 0.407, respectively. The same criteria for images of the second database for images with dimensions of 32 × 32 and 256 × 256 are 2.26–0.0046 (s) and 0.98–0.607 (s), respectively. Although the image is 8-times larger, the time spent has not changed much. On the other hand, the variance for applying different levels of noise such as fogging and hazing indicates the low computational complexity of the method. Another reason is in calculations, there are several stages of enhance. Therefore, compared to similar solutions, this time consuming is satisfactory. The lightweight structure used for the quality improvement phase with super-resolution is also acceptable comparing to similar methods. Some deep learning-based practices require considerable time to minimize the loss function. The proposed method takes less time than SRCNN, FSRCNN, A + , FSRCNN, and RFL methods. In the case of the BiCubic method, the computational rate is better for the first dataset, but this method has a higher relative error in improving the resolution. It should be noted that the proposed method is based on a deep structure. Figures 7 and 8, show a comparison of time complexity for different artifact enhancement levels and other methods in improving resolution, respectively.

Fig. 7
figure 7

Addition of noise and estimation of the computational complexity of the method of improving the quality of underwater images

Full size image
Fig. 8
figure 8

Calculating the average time of different Super-resolution methods compared to the proposed method indicates the algorithm's ability to reduce computational complexity

Full size image

Figures 8 and 9 also show the competitive complexity comparison for low/high dimensional optimized images among different methods in comparing with the proposed method in Fig. 10.

Fig. 9
figure 9

Calculating the average time of different Super-resolution methods compared to the proposed method in low dimensions of images

Full size image
Fig. 10
figure 10

Calculating the average time of different Super-resolution methods compared to the proposed method in high dimensions of images

Full size image

4.5 Discussion

The improvement set of the first part in Fig. 11 has been compared to other analogous methods. In this figure, the performance of the proposed algorithm in a sample of underwater natural sample is compared with similar strategies.

Fig. 11
figure 11

A comparison between similar methods and the proposed algorithm to improve the quality of underwater images

Full size image

The output of each image represents its satisfactory performance compared to other similar methods. The original underwater images arranged in the first column from left to right with artifacts. In second column, He et al. [50], Third column, Drews et al. [52], Fourth column, Tarel et al. [53], Fifth column, Ancuti et al. [54], sixth column, Barbosa et al. [55], Moghimi et al. [19], and last column proposed are displayed, respectively.

In the set shown in Fig. 12, the average loss function for all four scales of the image change is shown, converges on the optimal minimum values after 50 epochs. There are two main reasons for early emission and convergence to the minimum value. First, the proposed super-resolution algorithm is successful due to the lightweight structure and low layers and the convolutional mapping is removed from the processing set.

Fig. 12
figure 12

Loss function estimation for four different scales of images. First row from left to right show 2 × and 3 × loss functions, respectively. Also in the second row from left to right show 4 × and 8 × loss functions, respectively

Full size image

The performance accuracy of a super-resolution depends on the number of layers so that more layers can slightly enhance accuracy, but a trade-off is required for the right amount of time and specific repetitions.

That is, if the number of layers is increased to a certain level, then the convergence accuracy and, therefore, super-resolution at each scale is improved, but it will take a longer time, which is not desirable. In Figs, 13, 14, and 15, MSE, PSNR and SSIM measures are computed for the two datasets used, which are derived from low-artifact, medium-artifact and high-artifact modes, respectively. In the same vein, in Figs. 16, 17, 18, all three measures are calculated for the super-resolution in the network with low, medium, and high number of layers, with the mean layers producing better outputs.

Fig. 13
figure 13

Applying the proposed quality improvement algorithm to the images and calculating the mean MSE for both datasets by considering different low artifact, medium artifact, and high artifact modes

Full size image
Fig. 14
figure 14

Applying the proposed quality improvement algorithm to the images and calculating the mean PSNR for both datasets by considering different low artifact, medium artifact, and high artifact modes

Full size image
Fig. 15
figure 15

Applying the proposed quality improvement algorithm to the images and calculating the mean SSIM for both datasets by considering different low artifact, medium artifact, and high artifact modes

Full size image
Fig. 16
figure 16

Applying the proposed super-resolution algorithm to the images and calculating the mean MSE value for both datasets by considering low, medium and high number of layers

Full size image
Fig. 17
figure 17

Applying the proposed super-resolution algorithm to the images and calculating the mean PSNR value for both datasets by considering low, medium and high number of layers

Full size image
Fig. 18
figure 18

Applying the proposed super-resolution algorithm to the images and calculating the mean SSIM value for both datasets by considering low, medium and high number of layers

Full size image

5 Conclusion

Due to the low quality of underwater images, an efficient model with qualitative enhancement capability has been proposed. In this study, a two-step approach is proposed. In the first step, artifacts such as darkening, haze, and fogging are removed along with color correction. At this point, color correction of the images together with the removal of the artifacts yields an optimal optimization model. The second step involves the resolution enhancement of the image, which is performed at various scales using the deep learning neural network method. The results of the present study are applicable to systems that require multiple simultaneous underwater imaging functions. The results can contribute to the improvement of imaging systems and affect their multiple capabilities. Due to the poor image quality, sometimes the operator analysis will be time consuming and laborious and, therefore, an effective model is required.

The remarkable point of the algorithm in responding is that it uses a lightweight structure due to applying an efficient algorithm for different levels of noise, as well as exploiting super-resolution of images. Therefore, there is no have high computational complexity. In the future, by optimizing the image quality in both steps, we intend to optimize results and separate image parts.

This method can be used in the real world and suitable for various underwater images.