Weak-light image enhancement method for power operation sites based on strong-light suppression

Abstract

The present invention provides a method for enhancing weak-light images at power operation sites based on strong-light suppression, comprising the following steps: Step 1: judging strong light based on a sliding window; Step 2: light effect decomposition based on a layer decomposition network and Zero‑DCE weak-light enhancement based on a kernel selection module; Step 3: for the input weak-light RGB image, after grayscale conversion, use a sliding window to search the image and judge the strong-light area. The present invention considers both the overall weak light of the operation site image and the presence of a strong light source in the weak light from the perspective of practical application of power operations, and designs a strong-light judgment method based on a sliding window. The sliding window is used to segment the image, and the brightness is judged based on the average value and average deviation of the deviation between the grayscale sub-image and the strong-light threshold to find strong light; by performing light effect decomposition on the RGB image with strong light, the light effect layer is removed to reduce the interference of the strong light effect on the weak-light image enhancement.

Description

Technical Field

[0001] The present invention relates to an image enhancement method, in particular to a method for enhancing weak-light images at an electric power operation site based on strong-light suppression, and belongs to the field of computer science and technology. Background Art

[0002] Image quality in real scenes is often affected by environmental factors such as weather and lighting, causing some image information to disappear in darkness, exhibiting low-light characteristics. Research on low-light image enhancement can be categorized into two main categories: traditional enhancement methods and deep learning-based enhancement methods. Deep learning-based low-light image enhancement methods can be approached from the perspectives of supervised learning and unsupervised learning. The more authoritative traditional low-light image enhancement methods include histogram equalization-based methods, gamma correction, and methods based on Retinex theory. Traditional algorithms have the advantages of fast processing speed and easy deployment, but lack reference to real-world lighting conditions, leading to problems such as noise retention or amplification, artifacts, and color deviation.

[0003] With the development of artificial intelligence technologies such as computer vision, intelligent video surveillance systems are widely used in power operation scenarios. Affected by weather and light, power operation sites suffer from insufficient lighting, uneven lighting, and backlighting, resulting in poor image visibility. Low-light images seriously affect the accuracy of target detection and worker behavior recognition tasks at the operation site. Skeletal keypoint detection of power workers in low-light images results in a large number of false detections and missed detections, which seriously affects the use of skeletal keypoints to monitor worker behavior. Comparing the grayscale histogram, the grayscale histogram of the low-light image shows an uneven pixel distribution, with the majority of pixels having lower grayscale values. Furthermore, due to the particularity of the power industry, strong light sources are often used to supplement illumination during operations at night. Although strong light sources enhance local brightness, they interfere with weak light enhancement, resulting in local overexposure. Therefore, a low-light image enhancement method for power operation sites based on strong light suppression is proposed. Summary of the Invention

[0004] In view of this, the present invention provides a method for enhancing weak-light images at power operation sites based on strong-light suppression to solve or alleviate the technical problems existing in the prior art and at least provide a beneficial option.

[0005] The technical solution of the embodiment of the present invention is implemented as follows: a method for enhancing weak-light images at power operation sites based on strong-light suppression comprises the following steps:

[0006] Step 1: Determine the glare based on the sliding window;

[0007] Step 2: Light effect decomposition based on layer decomposition network and Zero-DCE weak light enhancement based on kernel selection module;

[0008] Step 3: For the input low-light RGB image, perform grayscale conversion and then use a sliding window to search the image and determine the bright light area;

[0009] Step 4: Use the layer decomposition network to decompose the light effect of the strong light image and obtain the background layer after removing the light effect layer;

[0010] Step 5: Implement low-light enhancement for the background layer image based on the Zero-DCE network of the kernel selection module.

[0011] Further preferably, in step 1, the input low-light RGB image is first converted into a grayscale image, and then 1 / 15 of the pixel width of the grayscale image is used as the side length a of the square sliding window, and the window is slid starting from the upper left corner with a step length;

[0012] The grayscale value of each pixel in the grayscale subimage is x (i,j) (0≤i<a,0≤j<a), the range of x is 0-255. The grayscale average value of a large number of strong photon images is calculated by public expression, and the calculation formula is:

[0013]

[0014] Finally, the average value AVG and average deviation AD of the deviation between each grayscale sub-image and the bright light threshold θ are calculated, and the brightness parameter S is further calculated. When S>1 and AVG>0, it is determined that there is bright light in the image.

[0015] Further preferably, in step 1, the calculation formula of the average value of the deviation between the grayscale image and the bright light threshold θ is:

[0016]

[0017] The calculation formula for the average deviation of the grayscale image and the bright light threshold θ is:

[0018]

[0019] Where a is the side length of the sliding window, and H[i] represents the number of pixels with gray level i in the grayscale image. The calculation formula of the brightness parameter S is:

[0020]

[0021] Further preferably, in step 2, the RGB image judged to have strong light is input into the layer decomposition network, and the φ G ,φ L and φ R Three independent networks and unsupervised loss decompose it into light effect layer, shadow layer and reflection layer. The light effect decomposition results of RGB images are as follows:

[0022]

[0023] Where I is the input image, G = φ G (I) is the light effect layer, L = φ L (I) is the shadow layer, R = φ R (I) is a reflective layer, Represents element-wise multiplication.

[0024] Further preferably, in step 4, the layer decomposition network uses a series of unsupervised losses. In the initial stage of training, G and L use G i and L i For supervision, directly perform L1 loss:

[0025] L init =|GG i |1+|LL i |1

[0026] Among them G i It is a smooth image generated by performing a second-order Laplace filter on the input image. i It is a grayscale image generated by taking the maximum value of the three channels at each position of the input image.

[0027] Further preferably, in step 4, the gradient map of G has a short-tail distribution, that is, the G map is smooth, the gradient is small, and there is almost no large gradient, while the gradient map of J has a long-tail distribution. By taking advantage of this feature, a loss called Gradient Exclusion Loss is used to separate the G and J layers as far as possible in the gradient space. The Gradient Exclusion Loss loss is defined as follows:

[0028]

[0029] Among them, G ↓n and J ↓n Represents G and J after bilinear interpolation downsampling, parameters and is the normalization factor, ||·|| F is the Frobenius norm.

[0030] Further preferably, in step 4, the Frobenius norm is a matrix norm, which is defined as follows:

[0031]

[0032] Among them, A * represents the conjugate transpose of A, σ iare the singular values ​​of A; in order to minimize the color shift in the decomposition output and balance the intensity value ranges of the three color channels in the background image J, a color constancy loss is set:

[0033]

[0034] Among them, (c1,c2)∈{(r,g),(r,b),(g,b)} represents the combination of two color channels. For the decomposition task, it is also required that the predicted three layer combination can also restore the original input image, that is, set the reconstruction loss:

[0035]

[0036] Each unsupervised loss is multiplied by its respective weight.

[0037] Further preferably, in step 5, the low-light image is taken as input, the DCE-SKNet learning curve parameter map is used, and then the low-light image is pixel-level adjusted using the brightness enhancement curve, and the enhanced image is obtained after multiple iterations;

[0038] The formula for the brightness enhancement curve is:

[0039] LE n (x) = LE n-1 (x)+A n (x)LE n-1 (x)(1-LE n-1 (x))

[0040] Among them, x represents the image pixel coordinate, n represents the number of iterations, A n Represents the curve parameter map, which is equal to the input image size.

[0041] Further preferably, in step 5, in order to reduce the number of parameters and the amount of calculation, the ordinary convolution in DCE-Net is replaced by a depth-wise separable convolution. The depth-wise convolution block first performs convolution on each channel separately to extract the information of a single channel, and then expands or compresses the channel of the input feature map through a 1×1 point-by-point convolution block to obtain a feature map of the expected size.

[0042] By making full use of spatial features and using receptive fields of different scales for multi-scale fusion, noise interference can be reduced. After the seventh convolutional layer of DCE-Net, a three-branch kernel selection module is added to adaptively adjust the receptive field size, dynamically select the appropriate path, and reduce the impact of noise on low-light enhancement.

[0043] For the input feature map U, 3×3, 5×5, and 7×7 convolution kernels are used to process it to obtain U′, U″, and U″′, respectively. The 5×5 convolution kernel is composed of two 3×3 dilated convolution kernels, and then the three are added together to obtain To integrate the information of all branches. Embed global information s through global average pooling,

[0044]

[0045] H and W are the height and width of the feature map respectively, and s then passes through the fully connected layer to produce a compact feature map z∈R d×C :

[0046] z=F fc (s) = δ(B(W s ))

[0047] Among them, δ represents the ReLU activation function, B represents batch normalization processing, and W∈R d×C , the d value is controlled by the compression ratio r:

[0048]

[0049] Where L is the minimum value of d;

[0050] In order to obtain weights of different spatial scales and thus obtain weighted fusion information of different receptive fields, a softmax operation is performed on the vector z channel direction to obtain:

[0051]

[0052]

[0053]

[0054] Among them, A,B,C∈R C×d , A c ∈R 1×d represents the cth row of A, α c is the cth element of α, α is the weight vector of U′, and finally the feature maps processed by convolution kernels of different sizes are multiplied by their corresponding weight vectors to obtain the final output feature map:

[0055]

[0056] Where V=[V1,V2,…,V c ],V c ∈R H×W ;

[0057] In order to enable the network to complete training with zero reference information, a series of non-reference losses are adopted, including spatial consistency loss, exposure control loss, color constancy loss and brightness smoothness loss.

[0058] Further preferably, in step 5, the spatial consistency loss is:

[0059] In order to ensure that the difference between the value of a pixel and the value of its adjacent pixels before and after image enhancement does not change too much, the following error is set:

[0060]

[0061] Where K is the number of local regions, Ω(i) is the four adjacent domains (upper, lower, left, and right) centered on region i; I and Y are the average intensity values ​​of the local regions in the input low-light image and the enhanced image, respectively;

[0062] Loss of exposure control:

[0063] The exposure control loss represents the distance between the average intensity value and the ideal exposure value E, so that the image has a good exposure value after enhancement. The formula is as follows:

[0064]

[0065] Where Y is the average intensity value of the local area in the enhanced image, E represents the grayscale of the RGB color space under ideal conditions, which is set to 0.6, and M represents the number of non-overlapping regions of 16×16;

[0066] Loss of color constancy:

[0067] According to the gray world color constancy hypothesis, the color of each sensor channel is averaged to the entire image. The potential color deviation in the enhanced image is corrected through color constancy loss. An adjustment relationship is established between the RGB channels to ensure that the average value of the RGB channels of the image after enhancement is as close as possible. The formula is as follows:

[0068]

[0069] Among them, J p and J q denote the average intensity values ​​of channels p and q respectively, and (p,q) denotes a set of channels belonging to ε;

[0070] Lighting smoothing loss:

[0071] Add illumination smoothness loss in each curve parameter map to maintain the monotonic relationship between adjacent pixels, that is, to make the brightness change between adjacent pixels not so significant. The formula is as follows:

[0072]

[0073] Where N represents the number of iterations, Represents the curve parameter diagram of each channel, represents the horizontal gradient of the image, Represents the vertical gradient of the image, ξ represents the RGB color three channels;

[0074] The total loss of the Zero-DCE weak light enhancement network of the kernel selection module is the weighted sum of the above four losses:

[0075] L total =W1L spa +W2L exp +W3L col +W4L tvA

[0076] Among them, W1, W2, W3, and W4 are the weight values ​​of the four losses.

[0077] The embodiment of the present invention adopts the above technical solution, which has the following advantages:

[0078] First, from the perspective of practical applications in power operations, this paper considers both the overall low-light situation in the worksite image and the presence of strong light sources within the low-light situation. A sliding window-based strong light detection method is designed. This method uses a sliding window to segment the image and determines brightness based on the average and average deviation of the grayscale sub-images from the strong light threshold to identify strong light.

[0079] Second, the present invention uses a light effect decomposition method based on a layer decomposition network to decompose the light effect of RGB images with strong light, remove the light effect layer, and reduce the interference of strong light effects on weak light image enhancement;

[0080] 3. This paper constructs a Zero-DCE low-light enhancement network based on the kernel selection module. The number of parameters and floating-point operations are significantly reduced compared with the original Zero-DCE. The subjective visual quality and objective evaluation indicators of the enhanced image are better than other advanced methods.

[0081] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present invention will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0083] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0084] Hereinafter, only certain exemplary embodiments are briefly described. As will be appreciated by those skilled in the art, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and description are to be considered as illustrative in nature and not restrictive.

[0085] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0086] like Figure 1 As shown, an embodiment of the present invention provides a method for enhancing weak-light images at a power operation site based on strong-light suppression, comprising the following steps:

[0087] Step 1: Determine the glare based on the sliding window;

[0088] Step 2: Light effect decomposition based on layer decomposition network and Zero-DCE weak light enhancement based on kernel selection module;

[0089] Step 3: For the input low-light RGB image, perform grayscale conversion and then use a sliding window to search the image and determine the bright light area;

[0090] Step 4: Use the layer decomposition network to decompose the light effect of the strong light image and obtain the background layer after removing the light effect layer;

[0091] Step 5: Implement low-light enhancement for the background layer image based on the Zero-DCE network of the kernel selection module.

[0092] In one embodiment, in step 1, the input low-light RGB image is first converted into a grayscale image. Then, a square sliding window is formed with a side length a of 1 / 15 of the grayscale image pixel width, and the window is slid starting from the upper left corner with a step length of a.

[0093] The grayscale value of each pixel in the grayscale subimage is x (i,j) (0≤i<a,0≤j<a), the range of x is 0-255. The grayscale average value of a large number of strong photon images is calculated by public expression, and the calculation formula is:

[0094]

[0095] Finally, the average value AVG and average deviation AD of the deviation between each grayscale sub-image and the strong light threshold θ are calculated, and the brightness parameter S is further calculated. When S>1 and AVG>0, it is determined that there is strong light in the image. The present invention sets the strong light grayscale threshold as θ=190; the calculation formula for the average value of the deviation between the grayscale image and the strong light threshold θ is:

[0096]

[0097] The calculation formula for the average deviation of the grayscale image and the bright light threshold θ is:

[0098]

[0099] Where a is the side length of the sliding window, and H[i] represents the number of pixels with gray level i in the grayscale image. The calculation formula of the brightness parameter S is:

[0100]

[0101] In one embodiment, in step 2, the RGB image judged to have strong light is input into the layer decomposition network, and the φ G ,φ L and φ R Three independent networks and unsupervised loss decompose it into light effect layer, shadow layer and reflection layer. The light effect decomposition results of RGB images are as follows:

[0102]

[0103] Where I is the input image, G = φ G (I) is the light effect layer, L = φ L (I) is the shadow layer, R = φ R (I) is a reflective layer, Represents element-by-element multiplication. The goal of this invention is to remove the light effect layer G and obtain a background layer that is not affected by the light effect. Thereby achieving the effect of removing the strong light effect, performing weak light enhancement based on the background layer J, and reducing the interference of the strong light effect.

[0104] In one embodiment, in step 4, the layer decomposition network uses a series of unsupervised losses. In the initial stage of training, G and L use G i and L i For supervision, directly perform L1 loss:

[0105] L init =|GG i |1+|LL i |1

[0106] Among them G i It is a smooth image generated by performing a second-order Laplace filter on the input image. iIt is a grayscale image generated by taking the maximum value of the three channels at each position of the input image. The gradient map of G has a short-tail distribution, that is, the G map is smooth, the gradient is small, and there is almost no large gradient, while the gradient map of J has a long-tail distribution. By taking advantage of this feature, a loss called Gradient Exclusion Loss is used to separate the G and J layers as far as possible in the gradient space. The Gradient Exclusion Loss loss is defined as follows:

[0107]

[0108] Among them, G ↓n and J ↓n Represents G and J after bilinear interpolation downsampling, parameters and is the normalization factor, ||·|| F is the Frobenius norm, which is a matrix norm defined as follows:

[0109]

[0110] Among them, A * represents the conjugate transpose of A, σ i are the singular values ​​of A; in order to minimize the color shift in the decomposition output and balance the intensity value ranges of the three color channels in the background image J, a color constancy loss is set:

[0111]

[0112] Among them, (c1,c2)∈{(r,g),(r,b),(g,b)} represents the combination of two color channels. For the decomposition task, it is also required that the predicted three layer combination can also restore the original input image, that is, set the reconstruction loss:

[0113]

[0114] Each unsupervised loss is multiplied by its own weight, where λ init and λ excl Set to 1, λ recon Set to 0.1, λ cc Set to 0.5.

[0115] In one embodiment, in step five, low-light enhancement is affected by noise, which may cause loss of local information around the noise, resulting in image blur. The present invention proposes a Zero-DCE low-light enhancement network based on a kernel selection module. The network takes the low-light image as input, uses the DCE-SKNet learning curve parameter map, and then performs pixel-level adjustment on the low-light image using the brightness enhancement curve. After multiple iterations, the enhanced image is obtained.

[0116] The formula for the brightness enhancement curve is:

[0117] LE n (x) = LE n-1 (x)+A n (x)LE n-1 (x)(1-LE n-1 (x))

[0118] Among them, x represents the image pixel coordinate, n represents the number of iterations, A n represents a curve parameter map, which is equal to the size of the input image. According to the above formula, each pixel of the input image is assigned an optimal high-order curve, enabling dynamic brightness adjustment. To reduce the number of parameters and computational complexity, the ordinary convolution in DCE-Net is replaced with a depthwise separable convolution. The depthwise convolution block first performs a separate convolution on each channel to extract the information of a single channel. Then, a 1×1 point-by-point convolution block is used to expand or compress the channels of the input feature map to obtain a feature map of the desired size. Compared to the standard convolution kernel, the depthwise separable convolution kernel can significantly reduce the number of network parameters with almost no loss in network accuracy.

[0119] Making full use of spatial features and using receptive fields of different scales for multi-scale fusion can reduce noise interference. In terms of multi-scale feature fusion, most existing methods are based on feature pyramid structures, combining features by element-wise addition or concatenation. Although they can combine feature maps of different scales, they ignore the spatial and channel specificity of features at different scales. After the seventh convolutional layer of DCE-Net, a three-branch kernel selection module is added to adaptively adjust the receptive field size, dynamically select the appropriate path, and reduce the impact of noise on low-light enhancement.

[0120] For the input feature map U, 3×3, 5×5, and 7×7 convolution kernels are used to process it to obtain U′, U″, and U″′, respectively. The 5×5 convolution kernel is composed of two 3×3 dilated convolution kernels, and then the three are added together to obtain To integrate the information of all branches. Embed global information s through global average pooling,

[0121]

[0122] H and W are the height and width of the feature map respectively, and s then passes through the fully connected layer to produce a compact feature map z∈R d×C :

[0123] z=F fc (s) = δ(B(W s ))

[0124] Among them, δ represents the ReLU activation function, B represents batch normalization processing, and W∈R d×C , the d value is controlled by the compression ratio r:

[0125]

[0126] Wherein, L is the minimum value of d, generally L=32;

[0127] In order to obtain weights of different spatial scales and thus obtain weighted fusion information of different receptive fields, a softmax operation is performed on the vector z channel direction to obtain:

[0128]

[0129]

[0130]

[0131] Among them, A,B,C∈R C×d , A c ∈R 1×d represents the cth row of A, α c is the cth element of α, α is the weight vector of U′, and finally the feature maps processed by convolution kernels of different sizes are multiplied by their corresponding weight vectors to obtain the final output feature map:

[0132]

[0133] Where V=[V1,V2,…,V c ],V c ∈R H×W ;

[0134] In order to enable the network to complete training with zero reference information, a series of non-reference losses are adopted, including spatial consistency loss, exposure control loss, color constancy loss and brightness smoothness loss.

[0135] Spatial consistency loss:

[0136] In order to ensure that the difference between the value of a pixel and the value of its adjacent pixels before and after image enhancement does not change too much, the following error is set:

[0137]

[0138] Where K is the number of local regions, Ω(i) is the four adjacent domains (upper, lower, left, and right) centered on region i; I and Y are the average intensity values ​​of the local regions in the input low-light image and the enhanced image, respectively. In this paper, the size of the local region is set to 4×4;

[0139] Loss of exposure control:

[0140] The exposure control loss represents the distance between the average intensity value and the ideal exposure value E, so that the image has a good exposure value after enhancement. The formula is as follows:

[0141]

[0142] Where Y is the average intensity value of the local area in the enhanced image, E represents the grayscale of the RGB color space under ideal conditions, which is set to 0.6, and M represents the number of non-overlapping regions of 16×16;

[0143] Loss of color constancy:

[0144] According to the gray world color constancy hypothesis, the color of each sensor channel is averaged to the entire image. The potential color deviation in the enhanced image is corrected through color constancy loss. An adjustment relationship is established between the RGB channels to ensure that the average value of the RGB channels of the image after enhancement is as close as possible. The formula is as follows:

[0145]

[0146] Among them, J p and J q denote the average intensity values ​​of channels p and q respectively, and (p,q) denotes a set of channels belonging to ε;

[0147] Lighting smoothing loss:

[0148] Add illumination smoothness loss in each curve parameter map to maintain the monotonic relationship between adjacent pixels, that is, to make the brightness change between adjacent pixels not so significant. The formula is as follows:

[0149]

[0150] Where N represents the number of iterations, Represents the curve parameter diagram of each channel, represents the horizontal gradient of the image, Represents the vertical gradient of the image, ξ represents the RGB color three channels;

[0151] The total loss of the Zero-DCE weak light enhancement network of the kernel selection module is the weighted sum of the above four losses:

[0152] L total =W1L spa +W2L exp +W3L col +W4L tvA

[0153] Among them, W1, W2, W3, and W4 are the weight values ​​of the four losses, which are 1, 10, 5, and 1600 respectively.

[0154] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various modifications and substitutions within the technical scope disclosed in the present invention, and such modifications and substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for enhancing weak-light images at power operation sites based on strong-light suppression, characterized in that: The following steps are involved: Step 1: Determine the glare based on the sliding window; Step 2: Light effect decomposition based on layer decomposition network and Zero-DCE weak light enhancement based on kernel selection module; Step 3: For the input low-light RGB image, perform grayscale conversion and then use a sliding window to search the image and determine the bright light area; Step 4: Use the layer decomposition network to decompose the light effect of the strong light image and obtain the background layer after removing the light effect layer; Step 5: Implement low-light enhancement on the background layer image using the Zero-DCE network based on the kernel selection module; In step 5, the low-light image is used as input, the DCE-SKNet learning curve parameter map is used, and then the low-light image is pixel-level adjusted using the brightness enhancement curve, and the enhanced image is obtained after multiple iterations; The formula for the brightness enhancement curve is: ； in, represents the image pixel coordinates, represents the number of iterations, Represents the curve parameter map, which is equal to the input image size; In step 5, the ordinary convolution in DCE-Net is replaced with a depth-wise separable convolution. The depth-wise convolution block first performs convolution on each channel separately to extract the information of a single channel, and then expands or compresses the channel of the input feature map through a 1×1 point-by-point convolution block to obtain a feature map of the expected size. We fully utilize spatial features and use receptive fields of different scales for multi-scale fusion. After the seventh convolutional layer of DCE-Net, we add a three-branch kernel selection module to adaptively adjust the receptive field size, dynamically select the appropriate path, and reduce the impact of noise on low-light enhancement. For the input feature map , respectively, using 3×3, 5×5, and 7×7 convolution kernels to process 、 and , where the 5×5 convolution kernel is composed of two 3×3 dilated convolutions, and then the three are added together to get To integrate information from all branches; Embedding global information through global average pooling , ； and are the height and width of the feature map, respectively. Then a compact feature map is generated through the fully connected layer : ； in, express activation function, represents batch normalization processing, , Compression Ratio To control: ； in, for The minimum value of In order to obtain the weights of different spatial scales and thus obtain the weighted fusion information of different receptive fields, the vector In the channel direction The operation results in: ； in, , express No. OK, for No. elements, for Finally, the feature maps processed by convolution kernels of different sizes are multiplied by their corresponding weight vectors to obtain the final output feature map: ； in, , ; In order to enable the network to complete training with zero reference information, a series of non-reference losses are adopted, including spatial consistency loss, exposure control loss, color constancy loss and brightness smoothness loss.

2. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 1, characterized in that: In step 1, the input low-light RGB image is first converted into a grayscale image, and then 1 / 15 of the grayscale image pixel width is used as the side length of the square sliding window. , starting from the upper left corner Slide for the step size; The grayscale value of each pixel in the grayscale subimage is , , The range is 0-255; the grayscale average value of a large number of strong photon images is calculated by public expression, and the calculation formula is: ； Finally, calculate the difference between each grayscale sub-image and the bright light threshold The mean of the deviations and mean deviation , further calculate the brightness parameters ,when , it is determined that there is strong light in the image.

3. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 2, characterized in that: In step 1, the grayscale image is compared with the bright light threshold The formula for calculating the mean value of the deviation is: ； Grayscale image and bright light threshold The formula for calculating the mean deviation is: ； in is the side length of the sliding window, Represents the gray level The number of pixels in the grayscale image; brightness parameter The calculation formula is: 。 4. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 1 is characterized by: In step 2, the RGB image that is judged to have strong light is input into the decomposition network layer, and the , and Three independent networks and unsupervised loss decompose it into light effect layer, shadow layer and reflection layer. The light effect decomposition results of RGB images are as follows: ； in is the input image, For the light effect layer, For the shadow layer, is the reflective layer, Represents element-wise multiplication.

5. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 1, characterized in that: In step 4, the layer decomposition network uses a series of unsupervised losses. In the initial stage of training, and use and To supervise and do it directly loss: ； in It is a smooth image generated by performing a second-order Laplace filter on the input image. It is a grayscale image generated by taking the maximum value of the three channels at each position of the input image.

6. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 5, characterized in that: In step four, The gradient map of has a short-tail distribution, while The gradient map has a long-tail distribution. By taking advantage of this feature, a loss called Gradient Exclusion Loss is used to separate the gradients as far as possible in the gradient space. and For two layers, the Gradient Exclusion Loss is defined as follows: ； in, and Represents the bilinear interpolation downsampling and ,parameter and is the normalization factor, is the Frobenius norm.

7. The method for enhancing weak-light images at power operation sites based on strong-light suppression according to claim 1, characterized in that: In step 5, spatial consistency loss: In order to ensure that the difference between the value of a pixel and the value of its adjacent pixels before and after image enhancement does not change too much, the following error is set: ； in, is the number of local regions, It is based on the region The four adjacent domains above, below, left and right of the center; and are the average intensity values ​​of the local areas in the input low-light image and the enhanced image respectively; Loss of exposure control: Exposure control loss indicates the difference between the average intensity value and the ideal exposure value. The distance between them makes the image have a good exposure value after enhancement. The formula is as follows: ； in, To enhance the average intensity value of the local area in the image, Indicates the grayscale level of the RGB color space in an ideal state, set to 0.6, Indicates the number of non-overlapping regions of 16×16; Loss of color constancy: According to the gray world color constancy hypothesis, the color of each sensor channel is averaged to the entire image. The potential color deviation in the enhanced image is corrected through color constancy loss. An adjustment relationship is established between the RGB channels to ensure that the average value of the RGB channels of the image after enhancement is as close as possible. The formula is as follows: ； in, and Represents channels and The average intensity value of Indicates belonging A set of channels; Lighting smoothing loss: Add illumination smoothing loss to each curve parameter map to maintain the monotonic relationship between adjacent pixels, that is, to make the brightness change between adjacent pixels not so significant. The formula is as follows: ； in, represents the number of iterations, Represents the curve parameter diagram of each channel, represents the horizontal gradient of the image, represents the vertical gradient of the image, Represents RGB color three channels; The total loss of the Zero-DCE weak light enhancement network of the kernel selection module is the weighted sum of the above four losses: ； in, are the weight values ​​of the four losses.

Images