Restoration and Enhancement of Underwater 1 Under-Exposure Images with Detail-Preserving

: Underwater images have great practical value in many fields such as underwater 7 archeology, seabed mining, and underwater exploration. Due to the complex underwater 8 environment, there are problems such as poor light, low contrast, and color degradation. 9 Traditional underwater image processing methods cannot well achieve the goal of clear display 10 under extreme conditions. This paper proposes a method for restoration and enhancement of 11 underwater under-exposure images that protects edge details and enhances image color. Firstly, 12 the underwater image was preprocessed, denoising with improved wavelet threshold function, 13 defogging with the Multi-Scale Retinex Color Restoration (MSRCR) and guided filter method. 14 Then, the method of adaptive exposure graph is used to enhance the under-exposure image. 15 Finally, the deep learning algorithm combined with the Non-Subsampled Contour Transform 16 (NSCT) technology is used to solve the problem of color degradation and edge texture weakening. 17 Experiments show that compared with other underwater image processing methods, this method 18 greatly improves the clarity of the image, enhances the color saturation and the edge texture details 19 of the image, and has a better visual effect.


23
At present, more and more activities are carried out underwater by human beings, such as

114
The task goal of super resolution is to convert the input low-resolution image into 115 high-resolution image, which is consistent with image de-noising and image de-blurring.

116
Super-resolution focuses on how images from small to large sizes are filled with new pixels; Image 117 de-noising is concerned with replacing the "contaminated" pixels with the correct ones without 118 changing the image size.

119
SRCNN is the first end-to-end super-resolution algorithm using CNN architecture (that is, 120 based on deep learning), which is better than the traditional multi-module integration method.

121
1)The structure of SRCNN is relatively simple. The whole convolutional network consists of 122 three convolutional layers, even without pooling and full connection layers; 123 2) Convolution operation is performed on low-resolution graphs to generate n1-dimensional 124 feature maps;

125
3) Conduct convolution operation on n1-dimensional feature map to generate n2 dimensional 126 feature maps;

127
4) The n2 dimension feature maps are convolved to generate super-resolved images.

128
There are three processing processes:

159
Therefore, it is more appropriate to take 3 as the general scale number.

161
The core of the Non-Subsampled Contour Transform (NSCT) transformation is Contourlet, the 162 transformation of the edge. Non-downsampling is based on the frequency domain, that is, for an 163 image, a frequency threshold is set first, and then the image is screened out with a filter that is greater 164 than or equal to the threshold frequency (of course, this is not a one-time screening process, but an iterative process of using a two-channel bandpass filter without downsampling).  The curve comparison chart of the hard threshold function, soft threshold function and the 187 threshold function used in this paper is shown in Figure 1. Comprehensive analysis shows that the 188 threshold function in this paper is better than the hard and soft threshold functions in processing.

196
Density function: Threshold calculation expression: Among them, γ 2 represents the variance of noise, γ x represents the standard deviation of the 202 sub-band coefficients, and j represents a certain layer of the layer.

203
For the calculation of γ 2 in (11), the method mentioned in [9] is used: In formula (13), n represents the length of the decomposition wavelet coefficient of each layer, 207 which is known from γ s 2 = γ 2 + γ x 2 : 209 Therefore, according to (12) (13) (14), the Bayesian threshold is jointly obtained, that is, the 210 adaptiveness of the Bayesian threshold between different layers is obtained, thereby highlighting 211 the benefits of the algorithm in this paper.

212
For the parameters δ and α in the improved threshold function, this paper uses the particle 213 swarm optimization algorithm mentioned in [10] to solve.

214
By applying the improved wavelet threshold function to the de-noising problem of (a) Original image (b) Denoised image 228

276
The optimization solution is divided into two steps. The first is to solve f(x) when φ( * ) is 277 ignored, which is an auto-closed value; the second is to introduce the guided filter into the solution,

294
Step 1: channel processing. Cut the initial image into three channels: R, G and B, and each 295 channel can get its own information. The formula is as follows:

297
In equation (22), i represents three channels of R, G, and B, and Y represents the original 298 image of CNN.

299
Step 2: CNN training. The formula is as follows:

304
Step 3: Image fusion. The formula is as follows:

306
Step 4: NSCT algorithm processing. The NSCT operation is performed on the fused image.

307
This method uses a tower decomposition algorithm to decompose the initial image into two parts,   image and the lack of clarity. The method used in this paper is better than other methods in 343 denoising, fogging, preserving details and protecting colors of underwater under-exposed images.

344
In order to further illustrate the clarity comparison of different methods, signal to noise ratio (SNR) 345 is adopted, as shown in table 1. The second stage. The method in this paper will be compared with the method proposed by D. analyze the corresponding implementation results of different methods in Fig.6, as shown in Table   364 2. Among them, the entropy of the image can be used to represent the statistical characteristics of becomes larger, the blurriness of the image will become smaller, that is, the more the image will be 369 the clearer.    improved. In Fig. 5