An object salvage method for turbid water based on artificial intelligence

Abstract

The present invention discloses a target salvage method for turbid water bodies based on artificial intelligence, belonging to the technical field of underwater search, and includes the following steps: S1. Use the side-scan sonar of an unmanned ship to scan the bottom of the water area where the salvage target is located to obtain side-scan sonar data; S2. Generate a sonar image; S3. Obtain a suspected salvage target; S4. Determine the circumferential-scan sonar data according to the suspected salvage target; S5. Determine the suspected target material of the suspected salvage target; S6. Judge whether the suspected target material is consistent with the salvage target material. If so, use an underwater robot to collect and transmit an image, and perform enhancement processing on the transmitted image to obtain a complete enhanced image for salvage. Otherwise, return to S1. By carrying a high-definition camera on the underwater robot and cooperating with an underwater enhancement algorithm, the present invention can not only make the image of the turbid bottom easy to see, but also improve the safety of the underwater robot operating in turbid waters and the salvage operation efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater search, and particularly relates to a target salvage method for turbid water based on artificial intelligence. Background Art

[0002] In the salvage industry, the positioning of the salvage target is the basis for operation. Traditional divers and underwater robots rely on optical vision search. The images transmitted back by the underwater robot are searched for the salvage target through image enhancement algorithms, target recognition models, and visual observation.

[0003] In recent years, water pollution incidents have occurred frequently, especially algal blooms and red tides. Excessive reproduction of microorganisms will cause extremely poor water visibility. Even with high-intensity floodlights for illumination, it is impossible to see targets in the middle and long distances. In waters with a large sediment content in rivers, the underwater visibility is affected by suspended sediment, and light will be affected by the absorption, scattering, and reflection of suspended sediment, and the propagation distance is very limited. In these turbid waters, due to limited visibility, even if the underwater robot can be positioned and obstacle-avoided through sonar, the robot relying on optical images will be unable to perform search tasks, resulting in the inability to carry out salvage tasks.

[0004] Therefore, a new salvage method is needed that can quickly and accurately search for and locate the salvage target in turbid waters affected by suspended sediment or water pollution, while ensuring the safety of the underwater robot operating in turbid waters. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a target salvage method for turbid water based on artificial intelligence.

[0006] The technical solution of the present invention is: a target salvage method for turbid water based on artificial intelligence includes the following steps:

[0007] S1. Use the side-scan sonar of an unmanned ship to scan the bottom of the water area where the salvage target is located to obtain side-scan sonar data;

[0008] S2. Generate a sonar image according to the side-scan sonar data;

[0009] S3. Input the sonar image into an image recognition model to obtain a suspected salvage target;

[0010] S4. Determine the circumferential scan sonar data according to the suspected salvage target;

[0011] S5. Input the circumferential scan sonar data into an acoustic echo recognition model to determine the suspected target material of the suspected salvage target;

[0012] S6. Determine whether the suspected target material is the same as the salvaged target material. If so, use the underwater robot to collect and transmit back images, and perform enhancement processing on the transmitted back images to obtain complete enhanced images for salvage. Otherwise, return to S1.

[0013] Further, S5 includes the following sub-steps:

[0014] S51. Use a deep neural network to perform feature fusion on the spectrum, power spectrum, and Mel cepstral coefficients of the circumferential scanning sonar data;

[0015] S52. Display the spectrum, power spectrum, and Mel cepstral coefficients after feature fusion in the form of images, and perform deep learning using a convolutional neural network to obtain an acoustic echo recognition model;

[0016] S53. Use the acoustic echo recognition model to determine the suspected target material.

[0017] Further, in S52, the loss function of the acoustic echo recognition model has the following expression:

[0018] ;

[0019] In the formula, represents the number of training samples, represents the number of classifications, represents the predicted probability of the th training sample in the th category, represents the adjustment factor, represents whether it belongs to the th sample in the category label, represents the logarithmic function.

[0020] Further, S6 includes the following sub-steps:

[0021] S61. Determine whether the suspected target material is the same as the salvaged target material. If so, use the underwater robot to collect and transmit back images, and enter S62. Otherwise, return to S1;

[0022] S62. Perform channel separation on the transmitted back images to obtain several single-channel images, where the single-channel images include R-channel images, G-channel images, and B-channel images;

[0023] S63. After performing non-linear gray-scale enhancement on the single-channel images, divide the single-channel into several sub-blocks;

[0024] S64. Calculate the gray-scale probability distribution of each sub-block, and determine the gray-scale reconstruction function of each sub-block according to the gray-scale probability distribution of each sub-block;

[0025] S65. Expand the grayscale values of each sub-block according to the grayscale reconstruction function of each sub-block to obtain the contrast of each sub-block;

[0026] S66. Determine the enhanced single-channel image according to the contrast of each sub-block;

[0027] S67. Synthesize each single-channel image to generate a color image;

[0028] S68. Obtain the enhanced image according to each single-channel image;

[0029] S69. Perform sampling filtering on both the color image and the enhanced image to obtain the filtered information of the color image and the filtered information of the enhanced image;

[0030] S610. Extract the detailed texture information of the color image and the detailed texture information of the enhanced image;

[0031] S611. Process the filtered information of the color image and the filtered information of the enhanced image using the normalized weight to obtain the low-frequency component;

[0032] S612. Process the detailed texture information of the color image and the detailed texture information of the enhanced image using the normalized weight to obtain the high-frequency component;

[0033] S613. Perform an inverse operation on the low-frequency component and the high-frequency component and reconstruct them to obtain the complete enhanced image.

[0034] Further, in S64, the gray-scale probability distribution of the

[0035] th sub-block is calculated by the formula:

[0036] In the formula, represents the number of times each gray scale appears in the sub-block, represents the size of the sub-block;

[0037] In S64, the gray-scale reconstruction function of the

[0038] th sub-block is expressed as:

[0039] Further, in S65, the contrast of the

[0040] th sub-block is calculated by the formula:

[0041] In the formula, represents the gray-scale reconstruction function of the th sub-block, represents the number of times each gray level appears in the sub-block, represents the size of the sub-block.

[0042] Furthermore, in S66, the enhanced single-channel image has the following expression:

[0043] ;

[0044] ;

[0045] In the formula, represents the interpolation weight, represents the description of the contrast plus the relative coordinates in the sub-block, represents the abscissa of the relative coordinates, represents the ordinate of the relative coordinates, represents the back-projected image, represents the abscissa of the pixel point of the back-projected image, represents the ordinate of the pixel point of the back-projected image, represents the abscissa of the centers of the four adjacent sub-blocks of the pixel point of the back-projected image, represents the ordinate of the centers of the four adjacent sub-blocks of the pixel point of the back-projected image, represents a constant.

[0046] Furthermore, in S68, the enhanced image has the following expression:

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula, represents the back-projected image, represents the set of the top 5% of the gray values in the back-projected image sorted from small to large, represents the image obtained by analyzing the back-projected image using the minimum value window, represents taking the maximum value, represents an intermediate variable, represents an exponent, represents the variance operation, represents taking the maximum value of the set of the top 5% of the gray values in the back-projected image sorted from small to large.

[0051] Furthermore, in S69, the filtering information of the color image The expression is:

[0052] ;

[0053] In the formula, represents the synthesis of the enhanced image corresponding to the single-channel image, represents the basic Gaussian filtering convolution kernel, represents the abscissa of the pixel point of the back-propagated image, represents the ordinate of the pixel point of the back-propagated image, represents the relative abscissa in the basic Gaussian filtering convolution kernel, represents the relative ordinate in the basic Gaussian filtering convolution kernel.

[0054] Furthermore, in S610, the detailed texture information of the color image The calculation formula is:

[0055] ;

[0056] In the formula, represents the synthesis of the enhanced image corresponding to the single-channel image, represents the basic Gaussian filtering convolution kernel, represents the abscissa of the pixel point of the back-propagated image, represents the ordinate of the pixel point of the back-propagated image, represents the abscissa of the relative coordinate, represents the ordinate of the relative coordinate.

[0057] The beneficial effects of the present invention are:

[0058] (1) By using the sidescan sonar carried by the unmanned boat for large-scale search and using the image recognition algorithm to recognize the sonar image, the present invention can help the operator quickly find the location of the salvage target and significantly improve the salvage efficiency;

[0059] (2) By using the ring-scan sonar carried by the underwater robot and cooperating with the acoustic echo recognition model, the present invention can be used to remotely judge the suspected target material in turbid waters, improving the salvage efficiency while reducing the risk of the robot operating in turbid waters;

[0060] (3) By using the high-definition camera carried by the underwater robot and cooperating with the underwater enhancement algorithm, the present invention can not only make the image of the turbid bottom easy to see, which can improve the safety of the underwater robot operating in turbid waters and the salvage operation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flowchart of a target salvage method for turbid water bodies based on artificial intelligence. Detailed implementation manners

[0062] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0063] As Figure 1 shown, the present invention provides a target salvage method for turbid water based on artificial intelligence, including the following steps:

[0064] S1. Use the sidescan sonar of the unmanned ship to scan the bottom of the water area where the salvage target is located to obtain sidescan sonar data;

[0065] S2. Generate a sonar image according to the sidescan sonar data;

[0066] S3. Input the sonar image into an image recognition model to obtain a suspected salvage target;

[0067] S4. Determine the circumferential scan sonar data according to the suspected salvage target;

[0068] S5. Input the circumferential scan sonar data into an acoustic echo recognition model to determine the suspected target material of the suspected salvage target;

[0069] S6. Judge whether the suspected target material is the same as the salvage target material. If so, use the underwater robot to collect and transmit the image, and perform enhancement processing on the transmitted image to obtain a complete enhanced image, and then perform salvage. Otherwise, return to S1.

[0070] In order to obtain the clearest image with the sidescan sonar carried by the unmanned ship, it is necessary to perform random multi-point depth sounding on the water area first to obtain the average depth of the water area, and adjust the frequency of the sidescan sonar transducer to emit and receive sound waves according to the measured depth. The shallower the water depth, the higher the frequency, and the deeper the water depth, the lower the frequency. Then, according to the size of the water area and the efficiency requirements, set the repeat scan range to 50%.

[0071] To ensure the complete scanning of the water area, it is necessary to divide the salvage water area into grids on the computer. The unmanned ship passes through each grid in turn according to the parallel grid scanning method to ensure that there is no missed or missing scanned water area. Finally, according to the sidescan sonar parameters, set the speed of the unmanned ship to 2.5 m / s.

[0072] Through the above steps, the sidescan sonar data with the highest resolution can be obtained, and the repeat scan range can be flexibly set according to the efficiency requirements, and the area of missed or missing scanning can be avoided to the greatest extent.

[0073] In S3 - S4, the unmanned boat needs to remove the sidescan sonar to ensure sufficient space and meet the power and communication bandwidth requirements of the underwater robot. The unmanned boat is connected to the underwater robot through a cable and brings the underwater robot above the position of the suspected salvage target identified in the sonar image. The unmanned boat will serve as a communication relay for the operation and data transmission of the underwater robot, and carry the underwater robot above the suspected salvage target to prepare for the diving salvage operation. The robot turns on the circumferential scanning sonar and dives to near the suspected target. Obstacle avoidance is an important aspect to ensure the safety of the underwater robot during operation in turbid waters. The circumferential scanning sonar is not only a tool for the underwater robot to dive and avoid obstacles in turbid waters, but also an important means for the underwater robot to judge whether the material of the suspected target is consistent with the salvage target when the optical image is ineffective in turbid waters.

[0074] The robot turns on the circumferential scanning sonar and dives. The circumferential scanning sonar will perform a 360° rotational scan, with a horizontal beam angle of 2° and a vertical beam width of 25°. The minimum detection distance is 0.75m, and the maximum detection distance is 50m, which can well assist in obstacle avoidance and ensure the safety of the underwater robot during operation in turbid waters.

[0075] The circumferential scanning sonar needs to fix the scanning angle at the azimuth of the suspected target to collect the acoustic echo information of the suspected target. Through the above steps, the underwater robot can safely dive to near the suspected target and can obtain the echo information of the suspected target at a long distance, reducing the risk of the underwater robot moving in turbid waters while improving the efficiency of the salvage operation.

[0076] In the embodiment of the present invention, S5 includes the following sub - steps:

[0077] S51. Feature fusion is performed on the spectrum, power spectrum, and mel - cepstral coefficients of the circumferential scanning sonar data using a deep neural network;

[0078] S52. The spectrum, power spectrum, and mel - cepstral coefficients after feature fusion are displayed in the form of an image, and deep learning is performed using a convolutional neural network to obtain an acoustic echo recognition model;

[0079] S53. Determine the material of the suspected target using the acoustic echo recognition model.

[0080] In S52, with time as the horizontal axis (ms) and frequency as the vertical axis (Hz), and the parameter intensity as the value of the pixel point (dB), a two - dimensional image is generated. In S52, the loss function adopts the Focal Loss combined with the cross - entropy training criterion.

[0081] In the embodiment of the present invention, in S52, the loss function of the acoustic echo recognition model The expression is:

[0082] ;

[0083] In the formula, represents the number of training samples, represents the number of classifications, represents the th training sample's predicted probability for the th class, represents the adjustment factor, represents whether it belongs to the th sample in the class label, represents the logarithmic function.

[0084] In an embodiment of the present invention, S6 includes the following sub-steps:

[0085] S61. Determine whether the suspected target material is the same as the salvaged target material. If so, use the underwater robot to collect and transmit back the image, and proceed to S62; otherwise, return to S1.

[0086] S62. Perform channel separation on the transmitted-back image to obtain a number of single-channel images. Among them, the single-channel images include R-channel images, G-channel images, and B-channel images.

[0087] S63. After performing non-linear gray-scale enhancement on the single-channel images, divide the single-channel into several sub-blocks.

[0088] S64. Calculate the gray-scale probability distribution of each sub-block, and determine the gray-scale reconstruction function of each sub-block according to the gray-scale probability distribution of each sub-block.

[0089] S65. According to the gray-scale reconstruction function of each sub-block, perform expansion processing on the gray-scale values of the sub-blocks to obtain the contrast of each sub-block.

[0090] S66. Determine the enhanced single-channel image according to the contrast of each sub-block.

[0091] S67. Synthesize each single-channel image to generate a color image.

[0092] S68. Obtain the enhanced image according to each single-channel image.

[0093] S69. Perform sampling filtering on both the color image and the enhanced image to obtain the filtered information of the color image and the filtered information of the enhanced image.

[0094] S610. Extract the detailed texture information of the color image and the detailed texture information of the enhanced image.

[0095] S611. Process the filtered information of the color image and the filtered information of the enhanced image using the normalized weight to obtain the low-frequency component.

[0096] S612. Process the detailed texture information of the color image and the enhanced detailed texture information using the normalized weights to obtain the high-frequency component;

[0097] S613. Perform an inverse operation on the low-frequency component and the high-frequency component and reconstruct them to obtain the complete enhanced image.

[0098] The input features include Mel cepstral coefficients, spectrum, and power spectrum. Among them, the extraction method of Mel cepstral coefficients is as follows: frame the sonar echo signal, perform pre-emphasis and windowing processing, then perform Fourier transform to obtain its spectrum, then use a band-pass filter for filtering, take the logarithm of the output of the filter, and then perform discrete cosine transform to obtain Mel cepstral coefficients. The method for spectrum extraction is to perform windowing processing on the echo data and then perform discrete fast Fourier transform of several points. First, intercept the power spectrum on the frequency axis, intercept the frequency where the ability of the sonar echo signal is mainly concentrated, perform windowing summation and averaging processing on the intercepted frequency part to obtain the average power spectrum of the echo signal. Finally, perform deep learning on the fused features to obtain the sonar echo signal recognition model.

[0099] In the embodiment of the present invention, in S64, the gray probability distribution of the sub-block is calculated by the formula:

[0100] ;

[0101] In the formula, represents the number of times each gray level appears in the sub-block, represents the size of the sub-block;

[0102] In S64, the gray reconstruction function of the sub-block is expressed as:

[0103] .

[0104] In the embodiment of the present invention, in S65, the contrast of the sub-block is calculated by the formula:

[0105] ;

[0106] In the formula, represents the gray reconstruction function of the sub-block, represents the number of times each gray level appears in the sub-block,

[0107] represents the size of the sub-block. In the embodiment of the present invention, in S66, the enhanced single-channel image The expression is:

[0108] ;

[0109] ;

[0110] In the formula, represents the interpolation weight, represents the description of the relative coordinates in the contrast addition sub-block, represents the abscissa of the relative coordinates, represents the ordinate of the relative coordinates, represents the back-propagated image, represents the abscissa of the pixel point of the back-propagated image, represents the ordinate of the pixel point of the back-propagated image, represents the abscissa of the center of the four adjacent sub-blocks of the pixel point of the back-propagated image, represents the ordinate of the center of the four adjacent sub-blocks of the pixel point of the back-propagated image, represents a constant.

[0111] In the embodiment of the present invention, in S68, the enhanced image The expression is:

[0112] ;

[0113] ;

[0114] ;

[0115] In the formula, represents the back-propagated image, represents the set of the top 5% of the gray values in the back-propagated image sorted from small to large, represents the image obtained by analyzing the back-propagated image using the minimum value window, represents taking the maximum value, represents an intermediate variable, represents an exponent, represents a variance operation, represents taking the maximum value of the set of the top 5% of the gray values in the back-propagated image sorted from small to large.

[0116] In the embodiment of the present invention, in S69, the filtering information of the color image The expression is:

[0117] ;

[0118] In the formula, represents the synthesis of the enhanced images corresponding to the single-channel images, represents the basic Gaussian filter convolution kernel, represents the abscissa of the pixel point of the back-propagated image, represents the ordinate of the pixel point of the back-propagated image, represents the relative abscissa in the basic Gaussian filter convolution kernel, represents the relative ordinate in the basic Gaussian filter convolution kernel.

[0119] In the embodiment of the present invention, in S610, the detailed texture information of the color image has the following calculation formula:

[0120] ;

[0121] In the formula, represents the synthesis of the enhanced image corresponding to the single-channel image, represents the basic Gaussian filter convolution kernel, represents the abscissa of the pixel point of the back-propagated image, represents the ordinate of the pixel point of the back-propagated image, represents the abscissa of the relative coordinate, represents the ordinate of the relative coordinate.

[0122] Those of ordinary skill in the art will realize that the embodiments described herein are to assist the reader in understanding the principles of the present invention and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations without departing from the essence of the present invention according to the technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. A target salvage method for turbid water bodies based on artificial intelligence, characterized in that, It includes the following steps: S1. Use the sidescan sonar of the unmanned ship to scan the bottom of the water area where the salvage target is located to obtain sidescan sonar data; S2. Generate a sonar image according to the sidescan sonar data; S3. Input the sonar image into the image recognition model to obtain a suspected salvage target; S4. Determine the circumferential scan sonar data according to the suspected salvage target; S5. Input the circumferential scan sonar data into the acoustic echo recognition model to determine the suspected target material of the suspected salvage target; S6. Judge whether the suspected target material is the same as the salvage target material. If so, use the underwater robot to collect and transmit the image, and perform enhancement processing on the transmitted image to obtain a complete enhanced image, and then carry out the salvage. Otherwise, return to S1; The S6 includes the following sub-steps: S61. Judge whether the suspected target material is the same as the salvage target material. If so, use the underwater robot to collect and transmit the image, and enter S62. Otherwise, return to S1; S62. Perform channel separation on the transmitted image to obtain several single-channel images. Among them, the single-channel images include R-channel images, G-channel images, and B-channel images; S63. After performing non-linear gray-scale enhancement on the single-channel images, divide the single-channel into several sub-blocks; S64. Calculate the gray-scale probability distribution of each sub-block, and determine the gray-scale reconstruction function of each sub-block according to the gray-scale probability distribution of each sub-block; S65. According to the gray-scale reconstruction function of each sub-block, perform expansion processing on the gray-scale value of the sub-block to obtain the contrast of each sub-block; S66. Determine the enhanced single-channel image according to the contrast of each sub-block; S67. Synthesize each single-channel image to generate a color image; S68. Obtain the enhanced image according to each single-channel image; S69. Perform sampling filtering on both the color image and the enhanced image to obtain the filtered information of the color image and the filtered information of the enhanced image; S610. Extract the detailed texture information of the color image and the detailed texture information of the enhanced image; S611. Use the normalized weight to process the filtered information of the color image and the filtered information of the enhanced image to obtain the low-frequency component; S612. Use the normalized weight to process the detailed texture information of the color image and the detailed texture information of the enhanced image to obtain the high-frequency component; S613. Perform inverse operation and reconstruction on the low-frequency component and the high-frequency component to obtain a complete enhanced image.

2. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that, The S5 includes the following sub-steps: S51. Use a deep neural network to perform feature fusion on the spectrum, power spectrum, and mel cepstral coefficients of the circumferential scan sonar data; S52. Display the spectrum, power spectrum, and mel cepstral coefficients after feature fusion in the form of an image, and perform deep learning using a convolutional neural network to obtain the acoustic echo recognition model; S53. Use the acoustic echo recognition model to determine the suspected target material.

3. The target salvage method for turbid water bodies based on artificial intelligence according to claim 2, characterized in that In the above S52, the loss function J of the acoustic echo recognition model Focal-CE has the following expression: where n represents the number of training samples, N represents the number of classifications, represents the predicted probability of the l-th training sample in the k-th category, γ represents the adjustment factor, represents the label indicating whether the l-th sample belongs to the k-th category, and log2(·) represents the logarithmic function.

4. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that, In the S64, the calculation formula for the gray-scale probability distribution p(k) of the k-th sub-block is: In the formula, h(k) represents the number of times each gray scale appears in the sub-block, and S represents the size of the sub-block; In the S64, the expression of the gray-scale reconstruction function C(k) of the k-th sub-block is:

5. The target salvage method for turbid water based on artificial intelligence according to claim 1, characterized in that In the S65, the calculation formula for the contrast T(k) of the k-th sub-block is: Wherein, C(k) represents the gray reconstruction function of the k-th sub-block, h(k) represents the number of occurrences of each gray level in the sub-block, and S represents the size of the sub-block.

6. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that, In the step S66, the expression of the enhanced single-channel image S(x, y) is: where ω i,j represents the interpolation weight, T i,j represents the description of the relative coordinates in the contrast addition sub-block, i represents the abscissa of the relative coordinates, j represents the ordinate of the relative coordinates, I(x, y) represents the back-projected image, x represents the abscissa of the back-projected image pixel point, y represents the ordinate of the back-projected image pixel point, represents the abscissa of the centers of the four adjacent sub-blocks of the back-projected image pixel point, represents the ordinate of the centers of the four adjacent sub-blocks of the back-projected image pixel point, and ∈ represents a constant.

7. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that In the step S68, the expression of the enhanced image M(x, y) is: B = Max(I min ); Wherein, I(x, y) represents the back-propagated image, and I min represents the set of the top 5% of the gray values in the back-propagated image sorted from smallest to largest, and I min (x, y) represents the image obtained by analyzing the back-propagated image using a minimum value window, Max(·) represents taking the maximum value, J(x, y) represents an intermediate variable, e represents the exponent, Var(·) represents the variance operation, and B represents taking the maximum value of the set of the top 5% of the gray values in the back-propagated image sorted from smallest to largest.

8. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that, In S69, the filtering information L of the color image F (x,y) is expressed as: Wherein, F(·) represents the synthesis of the enhanced image corresponding to the single-channel image, K(·) represents the basic Gaussian filter convolution kernel, x represents the abscissa of the pixel point of the back-propagated image, y represents the ordinate of the pixel point of the back-propagated image, m represents the relative abscissa in the basic Gaussian filter convolution kernel, and n represents the relative ordinate in the basic Gaussian filter convolution kernel.

9. The target salvage method for turbid water bodies based on artificial intelligence according to claim 1, characterized in that, In the S610, the detailed texture information H of the color image F (x,y) The calculation formula is: Wherein, F(·) represents the synthesis of the enhanced image corresponding to the single-channel image, K(·) represents the basic Gaussian filter convolution kernel, x represents the abscissa of the pixel point of the back-propagated image, y represents the ordinate of the pixel point of the back-propagated image, i represents the abscissa of the relative coordinate, and j represents the ordinate of the relative coordinate.

Images