Lightweight human skeleton key point detection method and device in weak light environment
Through improved cyclic generation adversarial network and axial decomposition integral regression bone key point detection network, the problem of detecting human bone key point in low-light environments is solved, and the detection effect of efficient and dependency-free pairing data sets is achieved.
Patent Information
- Application Number
- CN202411956588.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-29
- Publication Date
- 2025-04-25
AI Technical Summary
The prior art is difficult to effectively detect key points of human skeleton in low-light environments, and traditional methods rely on paired data sets and are difficult to apply in power operations.
The improved cyclic generation adversarial network (L2CycleGAN) was used for low-light image enhancement, and the human bone key points were extracted in combination with the axial decomposition integral regression bone key point detection network (ADIR-KNet).
Through unsupervised learning methods, get rid of the dependence on paired data sets, generate high-quality images, and improve the accuracy and efficiency of detection of key points in human bones.
Smart Images
Figure CN119887693A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a detection framework, and in particular to a method and a device for detecting lightweight human skeleton key points in a weak light environment. Background Art
[0002] Power operations involve dangerous operations such as strong electricity and high voltage. In view of the problem that illegal operations by power operators may cause serious accidents, it is urgent to improve the intelligent detection level of illegal operations by power operators to ensure the safe completion of power operations. With the construction of new power systems, intelligent equipment used to monitor the operation process or perform operations (mainly including cameras, mobile control balls and other intelligent terminals) has enabled massive power operation video data to be collected, providing a solid data foundation for the use of human posture estimation technology to realize the recognition of illegal behaviors in power operations. A reliable and efficient human skeleton key point extraction technology can improve the efficiency of identifying illegal behaviors of operators and reduce the risk of operating accidents.
[0003] With the construction of digital and intelligent power grids, intelligent monitoring equipment (such as cameras and mobile surveillance balls) is widely used, resulting in the collection of massive amounts of information on power operations. Human posture estimation technology can effectively monitor the posture and movements of operators based on the data recorded by monitoring equipment, such as images and videos, by identifying and analyzing a predefined set of human key points, so as to determine whether they have any illegal behaviors, which is of great significance for the safety monitoring of power operation sites. The task of detecting key points of the human skeleton is a prerequisite for achieving human posture estimation. Traditional methods for detecting key points of the human skeleton have large parameters, high computational complexity, and slow reasoning speed, which affects the efficiency of identifying illegal behaviors in power operations.
[0004] Existing low-light image enhancement methods often use supervised learning methods. During the network training process, low-light images and corresponding normal-light images need to be provided, which is heavily dependent on paired datasets. However, a large number of paired datasets are often difficult to obtain in power operation production activities. Unsupervised methods are used to convert low-light images to normal-light images, such as recurrent generative adversarial networks. However, due to their complex network models, they are prone to image information loss, resulting in low quality of generated images.
[0005] The traditional human skeleton key point detection network does not take into account the power operation scenes in low-light environments, such as cloudy days and nights. The existing low-light image enhancement methods (such as those based on Retinex theory) are mainly based on supervised learning methods, which heavily rely on paired datasets. However, low-light paired datasets for power operation sites are often difficult to collect.
[0006] The traditional heat map-based human skeleton key point detection network, such as HRNet, renders each key point in GroundTruth into a Gaussian heat map M i , in M i Each point in represents a skeleton key point K i The probability of falling at this position is calculated by calculating argmax(M j ) Get the coordinates L of the point with the highest probability in the heat map j , which is the coordinate position of the key point j. This method is used to calculate L j The entire heat map needs to be global Softmax normalized. Due to the large number of elements in the feature map, the probability distribution of the normalized feature map becomes flatter, and the distribution of each element value becomes denser and more continuous. When calculating expectations, it is highly dependent on the global probability, which will cause deviations and lead to performance degradation. Summary of the invention
[0007] In order to solve the defects in the prior art, the present invention discloses a lightweight human skeleton key point detection method in a weak light environment, and its technical solution is as follows:
[0008] A lightweight human skeleton key point detection method in a weak light environment, characterized by:
[0009] Step 1: determine whether to perform low-light image enhancement on the image obtained from the power operation site monitoring video. If so, input the image into the generator part of the trained low-light image enhancement cycle generative adversarial network L2CycleGAN, and then output the enhanced normal illumination image;
[0010] Step 2: Detect the key points of bones in the human body according to the image with normal illumination, generate a heat map, and input the heat map into the axial decomposition integral regression key point detection network ADIR-KNet to extract the coordinates of the key points of bones, thereby forming a human skeleton map and completing the task of detecting the key points of bones in the human body.
[0011] The present invention also discloses a lightweight human skeleton key point detection device in a weak light environment, which is characterized by:
[0012] Power operation site low-light image enhancement module: determines whether the image obtained from the power operation site monitoring video needs to be enhanced in low light conditions. If so, the image is input into the trained low-light image enhancement cyclic generative adversarial network L2CycleGAN, and then the enhanced normal-light image is output;
[0013] Human skeleton key point extraction module: Extract the skeleton key points in the human body based on the image with normal illumination, generate a heat map, and input the heat map into the axial decomposition integral regression skeleton key point detection network ADIR-KNet to extract the coordinates of the skeleton key points, thereby forming a human skeleton map and completing the human skeleton key point detection task.
[0014] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein when the program is running, the device where the non-volatile storage medium is located is controlled to execute the above-mentioned method.
[0015] The present invention also discloses an electronic device, characterized in that it comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the above method when they are executed.
[0016] Beneficial Effects
[0017] (1) The cyclic generative adversarial network structure is improved, and a detail generation module (DGM) is proposed. The DGM is applied to a novel network structure designed by the present invention. The entire network is used as a generator, and an identity consistency loss function is introduced to enhance the details of the generated image and improve the image quality. In addition, the dual-loop structure is used to get rid of the dependence of the low-light enhancement method on a large number of paired datasets.
[0018] (2) An axial decomposition integral regression module is proposed and applied to the HRNet network structure, getting rid of the dependence on global features. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a flow chart of the lightweight human skeleton key point detection method in a weak light environment;
[0020] Figure 2 This is a schematic diagram of the network structure of the L2CycleGAN proposed in the present invention;
[0021] Figure 3 A schematic diagram of a novel network structure proposed as a generator in the present invention;
[0022] Figure 4 This is a schematic diagram of the DGM module proposed in the present invention;
[0023] Figure 5 This is a schematic diagram of the ADIR-KNet structure. DETAILED DESCRIPTION
[0024] The present invention proposes a method for detecting key points of human skeleton in a weak light environment based on an improved cyclic generative adversarial network and an axial decomposition integral regression module. Through deep learning, local integral regression and other technologies, the brightness of low-light images is enhanced, thereby improving the detection accuracy of the lightweight human skeleton key point detection network. Through the improved double-loop structure, the dependence on paired data sets is eliminated, and images with better visual and objective evaluation are generated. The specific scheme is as follows:
[0025] The specific operation process of the lightweight human skeleton key point detection method in a weak light environment of the present invention is shown in Figure 1 ,This method divides the key point detection of human skeleton at the ,power operation site into the weak light image enhancement stage and ,the key point detection stage of human skeleton at the ,power operation site.
[0026] Step 1: It is necessary to determine whether the image obtained from the power operation site monitoring video needs to be enhanced for low-light images. If it is to be enhanced for low-light images, the image is input into the trained low-light image enhancement cycle generative adversarial network (L2CycleGAN), and then the enhanced normal illumination image is output;
[0027] This step is based on L2CycleGAN and uses an unsupervised learning method to get rid of the dependence on paired datasets and generate images with higher visual effects and objective evaluations. Figure 2 This is the network structure of L2CycleGAN proposed in this invention, where G and F are generators, D1 and D2 are discriminators, and L GAN To generate the adversarial loss, L CYC is the cycle consistency loss, L ICL is the identity consistency loss.
[0028] The working process of L2CycleGAN is as follows:
[0029] Step 1: Use the ImageNet standardization method to normalize the input image and adjust the pixel value range to [0, 1]. The image input size is processed into a 256×256 8-bit grayscale image through a solidified program. The brightness range of each pixel is 0 to 255, where 0 represents completely black and 255 represents completely white, to match the input requirements of the model;
[0030] Step 2: For the normalized image, perform low-light image determination. If it is a low-light image (average brightness is less than 50), it is input into the generator part of the trained L2CycleGAN. After the low-light image is input into the generator, an image with normal light can be obtained. Then, the image with normal light is output to the generator proposed in the present invention. Axial decomposition integral regression bone Skeleton key point detection network ( Axial Decomposition Integral Regression Keypoint Network, ADIR-KNet ) Otherwise, the image is directly input into the ADIR-KNet proposed in the present invention.
[0031] The training process of L2CycleGAN is as follows:
[0032] Step 1: Use the ImageNet normalization method to normalize the input image and adjust the pixel value range to [0, 1]. The image input size is processed to 256×256 through a Python script to match the input requirements of the model.
[0033] Step 2: Input the original low-light image into the generator G, which converts the low-light image into a high-brightness pseudo-enhanced image;
[0034] Step 3: Input the pseudo high-illumination image generated by generator G into generator F, and try to restore the pseudo high-illumination image to a reconstructed image consistent with the original image. In this process, calculate the cycle consistency loss L CYC To compare the reconstructed image with the original low-light image, ensure that the details of the original image are not lost during the enhancement process, L CYC The calculation formula is shown in (17):
[0035]
[0036] In the formula, G and F are generators, ||·||1 represents the L1 norm of the vector, Represent the sample expectations of the source domain and the target domain respectively, x and y represent the samples from the source domain and the target domain respectively, and P X and P Y Refers to the probability distribution of real samples in the source domain and target domain respectively;
[0037] Figure 3 A novel network structure proposed in the present invention includes a convolutional layer, a ReLU layer and other neural network structures, and introduces the detail enhancement module (Detail Generation Module, DGM) proposed in the present invention. The entire network is used as a generator to enhance the details of the generated image and improve the quality of the generated image.
[0038] The generator structure proposed in the present invention has the following working process:
[0039] (1) Preliminary feature extraction: The low-light image first passes through two convolutional layers for preliminary feature extraction.
[0040] (2) Detail feature extraction: After the initial feature extraction, the image is input into two DGM modules proposed by the present invention to further extract the detail features of the image. In DGM, features of different depths are fused through multiple jump connections, and two convolution kernels of different sizes are used to increase feature diversity. Its structure helps to optimize the generation of detail and color information of high-illuminance images. At the same time, ResnetBlock can deepen the network depth and give the generator more nonlinear mapping capabilities.
[0041] Figure 4 The DGM module proposed by the present invention has the following working process:
[0042] a. Let the input feature map be X n-1 , whose size is H×W×C in , where H is the height, W is the width, C in Assume W is the number of input channels. First, a multi-scale convolution operation is performed. In this module, there are three convolution kernels of different scales: 1×1, 3×3, and 5×5. s×s is the weight matrix of the s×s convolution kernel, S can be 1, 3, 5, and the size of the feature map obtained after s×s convolution is:
[0043] Y S×S =X n-1 *W S×S +b S×S (18)
[0044] where b S×S is the bias term of the convolution.
[0045] b. Perform feature connection. Let the feature map after connection be Z1, then Z1=[Y 1×1 ,Y 3×3 ,Y 5×5 ], where [·] indicates connection in the channel dimension; Z1 is then subjected to 1×1, 3×3, and 5×5 convolution and connection operations to obtain the feature map Z2, and the calculation process is similar.
[0046] c. After feature connection, feature adjustment is performed through 1×1 convolution to obtain feature map Z3.
[0047] d. Finally, after the Leaky ReLU activation function is processed, the output feature map X of the DGM module is obtained n , the calculation formula is as follows:
[0048]
[0049] (3) Further feature processing: After two convolutional layers, the features can be further extracted by 6 ResnetBlocks.
[0050] (4) Feature restoration and image output: The features are restored through two layers of deconvolution, and then the three channels of the image are output through one layer of convolution. Finally, the generated image is output after tanh activation.
[0051] Step 4: The pseudo high-illumination image generated by the generator G is input into the discriminator D2 and compared with the manually annotated high-illumination image. The discriminator D2 is used to distinguish between pseudo-enhanced images and real high-illumination images, and calculate the generation adversarial loss L GAN , used to optimize the ability of the generator G, L GAN The calculation process of is shown in formula (20):
[0052]
[0053] In the formula, D1 and D2 are discriminators.
[0054] Step 5: The manually annotated high-illumination image is input into the generator F, and the pseudo low-illumination image generated by F and the original low-illumination image are input into the discriminator D1 to generate the adversarial loss L GAN , used to optimize the capacity of F;
[0055] Step 6: Calculate L based on the manually annotated high-illumination image and the image reconstructed by F and G. CYC , to ensure that the generators G and F maintain the consistency of image content during the conversion process;
[0056] Step 7: Calculate the identity consistency loss L based on the original low-light image and the manually annotated high-light image ICL , the image is input into the discriminator corresponding to the generator. The smaller the value, the stronger the generator's ability to understand the structure of the input image. ICL The calculation process is shown in the formula:
[0057]
[0058] The L2CycleGAN network structure proposed in the present invention proposes a detail enhancement module (DGM) based on the cyclic generative adversarial network structure, and applies it to a novel network structure designed by the present invention. The entire network is used as a generator, and an identity consistency loss function is introduced to enhance the details of the generated image and improve the image quality. In addition, through the dual-loop structure, the low-light enhancement method is freed from the dependence on a large number of paired data sets.
[0059] Step 2: Extract the key points of bones in the human body according to the image with normal illumination, generate a heat map, and input the heat map into ADIR-KNet to extract the coordinates of the key points of bones, thereby forming a human skeleton map and completing the task of detecting the key points of bones in the human body.
[0060] The backbone structure of the human skeleton key point extraction network proposed in the present invention is based on the design concept of HRNet. Figure 5 As shown, the network backbone is divided into four stages, starting with high-resolution feature extraction, and gradually introducing low-resolution branches. Each branch runs in parallel and realizes the interaction and repeated fusion of features through the multi-scale fusion module, constantly exchanging information of different resolutions, strengthening the expression ability of multi-scale features, and thus improving the overall feature extraction effect. Finally, the features output by the backbone network are input into the axial decomposition integral regression module proposed in the present invention for further processing, and the specific steps are as follows:
[0061] Step 1 performs summation mapping operations along the x-axis and y-axis of the feature map output by the backbone network:
[0062]
[0063] Where M(i,j) is the value of the pixel in the feature map, H and W represent the height and width of the feature map respectively, and sum x (i) is a vector formed by compressing the values of the pixels in the i-th row of the feature map in the x-axis direction, sum y (j) is a vector formed by compressing the values of the pixels in the jth column of the feature map in the y-axis direction.
[0064] Step 2 performs Softmax normalization on the summed results, as shown below:
[0065] sum x =[sum x (0),sum x (1),...,sum x (W-1)] (24)
[0066] sum y =[sumy (0),sum y (1),...,sum y (H-1)] (25)
[0067] T x =Softmax(sum x ),T y =Softmax(sum y ) (26)
[0068] After Softmax normalization, we get the probability distribution T on the x-axis and y-axis. x and T y .
[0069] Step 3 Based on the calculated probability distribution T x and T y , calculate the expected values on the x-axis and y-axis respectively:
[0070]
[0071] The calculated E x and E y They represent the coordinates of the skeleton key points on the x-axis and y-axis respectively, that is, the key point detection result of the network proposed in the present invention is (E x ,E y ). If it is necessary to extract m bone key points in the human body, m heat maps will be generated. These m heat maps can be input into the human bone key point extraction network used in the present invention to extract the coordinates of m bone key points, and connect them according to the human bone structure. The scheme adopted by the present invention is as follows:
[0072] ●Head: nose tip, eyes, ears
[0073] ●Upper limbs: shoulders, elbows, wrists
[0074] ●Lower limbs: hip, knee, ankle
[0075] Finally, all key points and their connecting lines are drawn on the image, and the extracted key points are connected according to the human skeletal structure to form a complete human skeleton diagram.
[0076] The axial decomposition integral regression module proposed in the present invention gets rid of the dependence on global features. While inheriting the advantages of the traditional integral regression method in overcoming quantization errors, it can also avoid the problem of network performance degradation due to deviations. In addition, the axial decomposition integral regression mechanism effectively reduces the impact of the flatter probability distribution in the feature map after normalization, making the prediction more accurate.
[0077] The present invention proposes an axial decomposition-based integral regression module, which does not directly perform global normalization on the entire feature map, but sums along the x-axis and y-axis respectively, and then performs Softmax normalization calculation, thereby reducing the model's dependence on global features, reducing model deviation, and avoiding network performance degradation.
[0078] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.
Claims
1. A lightweight human skeleton key point detection method in a weak light environment, characterized by: Step 1: determine whether to perform low-light image enhancement on the image obtained from the power operation site monitoring video. If so, input the image into the generator part of the trained low-light image enhancement cycle generative adversarial network L2CycleGAN, and then output the enhanced normal illumination image; Step 2: Detect the key points of bones in the human body according to the image with normal illumination, generate a heat map, and input the heat map into the axial decomposition integral regression key point detection network ADIR-KNet to extract the coordinates of the key points of bones, thereby forming a human skeleton map and completing the task of detecting the key points of bones in the human body.
2. The method for detecting key points of lightweight human skeletons in a weak light environment according to claim 1 is characterized by: The working process of the low-light image enhancement cycle generation adversarial network L2CycleGAN is as follows: S1: The input image is normalized using the ImageNet standardization method to adjust the pixel value range to [0, 1]. The image input size is processed into a 256×256 8-bit grayscale image through a solidified program. The brightness range of each pixel is 0 to 255, where 0 represents completely black and 255 represents completely white, to match the input requirements of the model; S2: Perform low-light image judgment on the normalized image. If it is a low-light image, it is input into the generator part of the trained L2CycleGAN. After the low-light image is input into the generator, a normal-light image is obtained, and then the normal-light image is output to the axial decomposition integral regression skeleton key point detection network ADIR-KNet; otherwise, the image is directly input into ADIR-KNet.
3. According to the lightweight human skeleton key point detection method in a weak light environment according to claim 2, it is characterized by: the L2CycleGAN network structure and the specific steps of the training process are as follows: Step 1: Use the ImageNet normalization method to normalize the input image and adjust the pixel value range to [0, 1]. The image input size is processed to 256×256 through a Python script to match the input requirements of the model. Step 2: Input the original low-light image into the generator G, which converts the low-light image into a high-brightness pseudo-enhanced image; Step 3: Input the pseudo high-illumination image generated by generator G into generator F, and try to restore the pseudo high-illumination image to a reconstructed image consistent with the original image. In this process, calculate the cycle consistency loss L CYC To compare the reconstructed image with the original low-light image, ensure that the details of the original image are not lost during the enhancement process, L CYC The calculation formula is shown in (1): In the formula, G and F are generators, ||·||1 represents the L1 norm of the vector, Represent the sample expectations of the source domain and the target domain respectively, x and y represent the samples from the source domain and the target domain respectively, and P X and P Y Refers to the probability distribution of real samples in the source domain and target domain respectively; Step 4: The pseudo high-illumination image generated by the generator G is input into the discriminator D2 and compared with the manually annotated high-illumination image. The discriminator D2 is used to distinguish between pseudo-enhanced images and real high-illumination images, and calculate the generation adversarial loss L GAN , used to optimize the ability of the generator G, L GAN The calculation process of is shown in formula (2): In the formula, D1 and D2 are discriminators. Step 5: The manually annotated high-illumination image is input into the generator F. The pseudo low-illumination image generated by F and the original low-illumination image are input into the discriminator D1 to calculate the generated adversarial loss L GAN , used to optimize the capacity of F; Step 6: Calculate L based on the manually annotated high-illumination image and the image reconstructed by F and G. CYC , to ensure that the generators G and F maintain the consistency of image content during the conversion process; Step 7: Calculate the identity consistency loss L based on the original low-light image and the manually annotated high-light image ICL , the image is input into the discriminator corresponding to the generator. The smaller the value, the stronger the generator's ability to understand the structure of the input image. ICL The calculation process is shown in the formula:
4. The method for detecting key points of lightweight human skeleton in a weak light environment according to claim 1 is characterized by: The step 2 further includes the following contents: First, the convolution operation, downsampling and upsampling operations are described in detail as follows: It is agreed that H and W represent the height and width of the input feature map respectively. out , W out They represent the height and width of the output feature map, respectively. F is the convolution kernel size, the step size is S, and the number of padded pixels is P. a. After the convolution operation, the feature map size calculation formula is as follows: b. Downsampling operation, the specific method is average pooling, which can smooth the feature map, reduce the impact of noise, and reduce the amount of data and calculation. The formula for calculating the feature map size after average pooling is as follows: c. Upsampling operation, the specific method is the nearest neighbor interpolation method. Assuming the upsampling multiple is r, the formula for calculating the feature map size after upsampling is as follows: H out =H×r (8) IN out =W×r (9) The working process of ADIR-KNet is as follows: First, the image with normal illumination is input into the human skeleton key point detection network. The backbone network of the network is divided into four stages. The specific process is as follows: Step 1: stage 1 uses the image output from step 1 with an image size of 256×256 as the initial input of ADIR-KNet. After convolution and downsampling operations, the stage 2 input image is obtained with an image size of 64×64. At this stage, the convolution kernel size F=4, the step size S=4, and the padding P=0; Step 2: Receive the input image obtained from stage 1 after convolution and downsampling operations; in stage 2, extract and process the input image through a series of convolution, upsampling and downsampling operations. After processing, the output image continues to serve as the input of stage 3; in this stage, for the convolution operation, take the convolution kernel size F = 1, step size S = 1, and padding P = 0; for the downsampling operation, take the convolution kernel size F = 2, step size S = 2, and padding P = 0; for the upsampling operation, take the upsampling multiple r = 2; Step 3: Receive the 64×64 image processed by stage 2; in stage 3, there will also be convolution, upsampling and downsampling operations to further extract features; These operations help the network capture more abstract and representative features. The processed image size is 64×64 and then passed to stage4. The parameter configuration is the same as stage2. Step 4: Receive the 64×64 image from stage 3. In this stage, continue the feature extraction operation, using convolution, upsampling and downsampling to mine deep feature information. The parameter configuration is the same as stage 2. Finally, the feature map extracted by the backbone network is output to the axial decomposition integral regression module. The process is as follows: Step 1: Perform summation and mapping operations along the x-axis and y-axis of the feature map output by the backbone network: Where M(i,j) is the value of the pixel in the feature map, H and W represent the height and width of the feature map respectively, and sum x (i) is a vector formed by compressing the values of the pixels in the i-th row of the feature map in the x-axis direction, sum y (j) is a vector formed by compressing the value of the pixel point in the jth column of the feature map in the y-axis direction; Step 2: Perform Softmax normalization on the summed results, as shown below: sum x =[sum x (0),sum x (1),...,sum x (W-1)] (12) sum y =[sum y (0),sum y (1),...,sum y (H-1)] (13) <h2 style=";text-align:left;direction:ltr">T<h2 style=";text-align:left;direction:ltr"> x <h2 style=";text-align:left;direction:ltr"> =Softmax(sum<h2 style=";text-align:left;direction:ltr"> x <h2 style=";text-align:left;direction:ltr"> ),T<h2 style=";text-align:left;direction:ltr"> y <h2 style=";text-align:left;direction:ltr"> =Softmax(sum<h2 style=";text-align:left;direction:ltr"> y <h2 style=";text-align:left;direction:ltr"> ) (14) After Softmax normalization, we get the probability distribution T on the x-axis and y-axis. x and T y ; Step 3: According to the calculated probability distribution T x and T y , calculate the expected values on the x-axis and y-axis respectively: The calculated E x and E y They represent the coordinates of the skeleton key points on the x-axis and y-axis respectively, that is, the key point detection result of the network proposed in the present invention is (E x ,E y ).
5. A lightweight human skeleton key point detection device in a weak light environment, characterized by: Power operation site low-light image enhancement module: determines whether the image obtained from the power operation site monitoring video needs to be enhanced in low light conditions. If so, the image is input into the generator part of the trained low-light image enhancement cycle generative adversarial network L2CycleGAN, and then the enhanced normal-light image is output; Human skeleton key point extraction module: Extract the skeleton key points in the human body based on the image with normal illumination, generate a heat map, and input the heat map into the axial decomposition integral regression skeleton key point detection network ADIR-KNet to extract the coordinates of the skeleton key points, thereby forming a human skeleton map and completing the human skeleton key point detection task.
6. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored program, wherein the program controls the device where the non-volatile storage medium is located to execute the method according to any one of claims 1 to 4 when the program is executed.
7. An electronic device, characterized in that: It comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the method described in any one of claims 1 to 4 when executed.