Shadow elimination method for personnel image detection in power training field

A cycle-consistent adversarial network addresses the challenge of shadow interference in personnel image detection by iteratively refining images using a GAN, improving recognition accuracy and enabling better operator analysis in the electric power training field.

CN113780298BActive Publication Date: 2024-12-03STATE GRID SHANGHAI MUNICIPAL ELECTRIC POWER CO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111088066.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-16
Publication Date
2024-12-03
Estimated Expiration
2041-09-16

AI Technical Summary

Technical Problem

Existing image recognition algorithms in the electric power training field suffer from reduced recognition accuracy due to the presence of shadows and reflections, particularly in complex video scenes with frequent operator posture changes, necessitating improved shadow removal methods for enhanced personnel detection.

Method used

A cycle-consistent adversarial network is employed to remove shadows in personnel images, utilizing a generative adversarial network (GAN) with two generators and discriminators to iteratively refine image features without requiring pixel-level shadow identification, achieving shadow-free images through adversarial training and cycle consistency.

Benefits of technology

The method effectively reduces the influence of shadows, enhancing personnel recognition and enabling more accurate operator action and behavior analysis by generating high-quality shadow-free images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113780298B_ABST
    Figure CN113780298B_ABST
Patent Text Reader

Abstract

The present invention provides a method for removing shadows in the detection of images of people in an electric power training field, comprising: obtaining sample data; a discriminator determines whether the shadow image comes from a generator or a real shadow image, and feeds back the first error between the shadow image and the real shadow image to the generator, and the generator reduces the first error; the discriminator determines whether the shadow-free image comes from the generator or the real shadow-free image, and feeds back the second error between the shadow-free image and the real shadow-free image to the generator, and the generator reduces the second error; the generator and the discriminator perform repeated adversarial training to obtain an optimal shadow removal network model; the shadow removal network model is used to remove shadows, and an image of a person in an electric power training field with shadows is input, and an image of a person in an electric power training field without shadows is output. Through a large amount of training on the shadow data set, the neural network is constantly self-gamed in the confrontation, so as to better realize the style transfer of the shadow image.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for removing shadows from an image, in particular to a method for removing shadows in personnel image detection in an electric power training field. Background Art

[0002] Motion image understanding based on video analysis has always been a hot topic in the computer field. As early as 2005, someone in the field of production safety proposed a helmet wearing status detection algorithm in a paper: helmet detection is performed based on features based on gradient direction histograms, or the face area is first detected through skin color detection, and after obtaining the face image, the helmet is detected through feature extraction. In addition, some people have proposed compliance detection of personal protective equipment on construction sites, using computer vision methods to conduct safety inspections on site videos, or using deep learning YOLO-v3 networks to achieve helmet wearing detection.

[0003] In the field of power construction, existing technologies use image recognition methods to detect the identity of power construction personnel; there are also security deployment systems based on image recognition technology, which discuss real-time monitoring and alarming of negligence and dangerous behaviors during construction; or through video acquisition equipment, using gradient features and support vector machine methods to study the intrusion of personnel on the power operation site, etc. It can be seen that video analysis is widely used in the field of production safety, especially in power operation safety monitoring, and has achieved good results.

[0004] However, due to the complexity of the video scenes of power grid monitoring and the frequent changes in the postures of operators, the recognition accuracy of operators will be greatly reduced. In order to improve the level of safety supervision of substation on-site operations, the target recognition of on-site operators was studied based on the traditional image recognition algorithm, but the recognition rate of the algorithm decreased in the presence of human shadows and reflections of accumulated water.

[0005] Therefore, based on the image recognition algorithm, eliminating image shadows and improving the recognition rate has become an urgent problem to be solved. Summary of the invention

[0006] The present invention provides a method for removing shadows in personnel image detection in an electric power training field. The method proposed in the present invention can effectively reduce the influence of shadows in images on personnel recognition, thereby realizing personnel extraction and further realizing a higher level of operator action and behavior analysis.

[0007] To achieve the above-mentioned and other related purposes, the present invention provides a method for removing shadows in personnel image detection in an electric power training field, comprising the following steps:

[0008] S1. Acquire sample data, where the sample data includes a true shadow image A and a true shadow-free image B;

[0009] S2, conduct adversarial training of the generative adversarial network, the real shadow image A is used as the input data of the generator A2B, the generator A2B outputs the shadow-free image B1, the shadow-free image B1 is used as the input data of the generator B2A, and the generator B2A outputs the shadow image A1;

[0010] The real shadow-free image B is used as the input data of the generator B2A, and the generator B2A outputs the shadow image A2. The shadow image A2 is used as the input data of the generator A2B, and the generator A2B outputs the shadow-free image B2.

[0011] Discriminator D A Determine whether the shadow image A2 is generated by the generator B2A or the real shadow image A, and feed back the first error between the shadow image A2 and the real shadow image A to the generator B2A, so that the generator B2A reduces the first error and improves the authenticity of the generated image; the discriminator D B Determine whether the shadow-free image B1 comes from the generator A2B or the real shadow-free image B, and feed back a second error between the shadow-free image B1 and the real shadow-free image B to the generator A2B, so that the generator A2B reduces the second error and improves the authenticity of the generated image;

[0012] Generator A2B and generator B2A are respectively connected to the discriminator D B and the discriminator D A Conduct repeated adversarial training to obtain the optimal shadow removal network model;

[0013] S3. Use the shadow removal network model to remove shadows, input images of people in the electric power training field with shadows, and output images of people in the electric power training field without shadows.

[0014] Preferably, the generator A2B comprises an encoder, a converter and a decoder;

[0015] The encoder inputs image features extracted from the sample data by a convolutional neural network;

[0016] The converter combines similar features of the image features and determines, based on the similar features, how to convert a feature vector of the image from a feature vector of a real shadow image A data set into a feature vector of a shadow-free image B data set;

[0017] The decoder uses a deconvolution layer to restore low-level features from the feature vector, and converts the obtained low-level features to generate a generated image in the shadow-free image B dataset and the same size as the real shadow image A dataset.

[0018] Preferably, the first error and the second error are both calculated by generating an adversarial loss of a network, and the adversarial loss is applied to a first mapping function and a second mapping function, wherein the first mapping function represents a mapping of a dataset X of shadow images to a dataset Y of shadow-free images, and the second mapping function represents a mapping of a dataset Y of a real shadow-free image B to a dataset X of a real shadow image A, and the mapping target of the first mapping function is:

[0019]

[0020] Among them, y represents the sample in the data set Y, x represents the sample in the data set X, and D Y (y) indicates the shadow-free image in the discriminator D Y The closer the score is to 1, the more realistic the discriminator thinks the image is. G(x) is the sample generated by the generator based on x and has the same distribution as the dataset Y. Y (G(x)) is the score obtained by the discriminator based on the generated samples. If D Y The more fake the generated samples are, the more likely D Y Rating D Y The closer (G(x)) is to 0, the better the 1-D Y The closer (G(x)) is to 1, Represents the probability distribution P data(x) The average amount of information contained, Represents the probability distribution P data(y) The average amount of information contained.

[0021] Preferably, a cycle consistency loss function is used for excitation, and the cycle consistency loss function is:

[0022]

[0023] Preferably, in step S2, the discriminator D B When the probability of determining whether the shadow-free image B1 comes from the generator A2B or the real shadow-free image B is 0.5, the training of the shadow removal network model reaches the optimal level.

[0024] Preferably, step S1 also includes screening of sample data, specifically: deleting sample data with incomplete images, clarity below the normal recognition range of human eyes, and unqualified content.

[0025] Preferably, step S1 further includes processing the sample data, specifically: using an open CV: resize algorithm to crop or scale the image to a uniform size for all sample data, and using a Z-score standardization method to normalize the data set.

[0026] In summary, the present invention applies a cycle-consistent adversarial network to solve the shadow removal problem. Traditional shadow removal algorithms need to mark the shadow area of ​​the detection target during the shadow detection stage when performing shadow removal, and cannot accurately identify the shadow area. However, the Cycle GAN adversarial network used in the present invention does not need to mark the shadow target during the game between the generator and the discriminator, and does not need to identify whether each pixel has been converted between pixels. When Cycle GAN performs shadow removal, through a large amount of training on the shadow data set, the neural network constantly plays a self-game in the confrontation, thereby better realizing the style transfer of the shadow image, and under the premise of ensuring that the generator and the discriminator are cycle-consistent during the training process, the shadow removal is better achieved, and a good shadow removal effect is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic flow chart of a method for removing shadows in personnel image detection in an electric power training field provided by an embodiment of the present invention;

[0028] Figure 2 A flowchart of a method for removing shadows in personnel image detection in an electric power training field provided by an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of a generator in a method for removing shadows in personnel image detection in an electric power training field provided by an embodiment of the present invention;

[0030] Figure 4 A schematic diagram of a discriminator in a shadow elimination method for personnel image detection in an electric power training field provided by an embodiment of the present invention;

[0031] Figure 5 A diagram of the adversarial loss function of a shadow removal method in personnel image detection in an electric power training field provided by an embodiment of the present invention;

[0032] Figure 6 HSV histogram after the shadow and no-shadow images in the prior art are converted into HSV images. DETAILED DESCRIPTION

[0033] The following is combined with Figure 1- and specific implementation methods further describe in detail the shadow elimination method in the personnel image detection in the electric power training field proposed by the present invention. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the implementation method of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the drawings. It should be noted that the structure, proportion, size, etc. illustrated in the drawings of this specification are only used to match the content disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so it has no technical substantive significance. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effect that the present invention can produce and the purpose that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0034] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0035] See also Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a method for removing shadows in personnel image detection in an electric power training field, comprising:

[0036] S1. Acquire sample data, where the sample data includes a true shadow image A and a true shadow-free image B;

[0037] S2, conduct adversarial training of the generative adversarial network, the real shadow image A is used as the input data of the generator A2B, the generator A2B outputs the shadow-free image B1, the shadow-free image B1 is used as the input data of the generator B2A, and the generator B2A outputs the shadow image A1;

[0038] The real shadow-free image B is used as the input data of the generator B2A, and the generator B2A outputs the shadow image A2. The shadow image A2 is used as the input data of the generator A2B, and the generator A2B outputs the shadow-free image B2.

[0039] Discriminator DA Determine whether the shadow image A2 is generated by the generator B2A or the real shadow image A, and feed back the first error between the shadow image A2 and the real shadow image A to the generator B2A, so that the generator B2A reduces the first error and improves the authenticity of the generated image; the discriminator D B Determine whether the shadow-free image B1 comes from the generator A2B or the real shadow-free image B, and feed back a second error between the shadow-free image B1 and the real shadow-free image B to the generator A2B, so that the generator A2B reduces the second error and improves the authenticity of the generated image;

[0040] Generator A2B and generator B2A are respectively connected to the discriminator D B and the discriminator D A Conduct repeated adversarial training to obtain the optimal shadow removal network model;

[0041] S3. Use the shadow removal network model to remove shadows, input images of people in the electric power training field with shadows, and output images of people in the electric power training field without shadows.

[0042] In this embodiment, see Figure 3 , the generator A2B includes an encoder, a converter and a decoder. The encoder inputs the image features extracted from the sample data by the convolutional neural network; the converter combines the similar features of the image features, and based on the similar features, determines how to convert the feature vector of the image from the feature vector of the real shadow image A data set to the feature vector of the shadow-free image B data set; the decoder uses the deconvolution layer to restore the low-level features from the feature vector, and converts the obtained low-level features to generate a generated image in the shadow-free image B data set and the same size as the real shadow image A data set. It is used to implement the following three steps:

[0043] 1) Encoding

[0044] First, the convolutional neural network extracts features from the input image, that is, the sample data, and inputs the extracted features into the designed encoder to obtain an output of a certain size, for example, an output of size [64, 64, 256].

[0045] 2) Conversion

[0046] The converter includes several network layers, the function of which is to combine different similar features of the image, and then based on these features, determine how to convert the feature vector of the image from the feature vector of the real shadow image A dataset to the feature vector of the shadow-free image B dataset.

[0047] 3) Decoding

[0048] Decoding is a process that is completely opposite to encoding. The deconvolution layer is used to restore the low-level features from the feature vector. Finally, the obtained low-level features are converted to generate an image in the shadow-free image B dataset, and a generated image with the same size as the real shadow image A dataset is obtained.

[0049] The generator B2A is similar to the generator A2B, but implements the corresponding generation from the shadow-free image B dataset to the real shadow image A dataset.

[0050] In this embodiment, see Figure 4 The input of the discriminator is an original shadow image. The function of the discriminator is to continuously try to predict whether it is the original shadow image or the output image generated by the generator. In this embodiment, two generators and two discriminators are set up. The convolutional neural network of the discriminator is used to extract vital signs from the image, thereby generating a convolutional layer with a one-dimensional output.

[0051] In this embodiment, the adversarial loss is applied to a first mapping function and a second mapping function, wherein the first mapping function represents a mapping of a data set X of shadow images to a data set Y of shadow-free images, and the second mapping function represents a mapping of a data set Y of a true shadow-free image B to a data set X of a true shadow image A. Below, the first mapping function is denoted by G and the second mapping function is denoted by F.

[0052] For the mapping function G: X→Y and its discriminator D Y , the goal is expressed as:

[0053]

[0054] Among them, y represents the sample in the data set Y, x represents the sample in the data set X, and D Y (y) indicates the shadow-free image in the discriminator D Y The closer the score is to 1, the more realistic the discriminator thinks the image is. G(x) is the sample generated by the generator based on x and has the same distribution as the dataset Y. Y (G(x)) is the score obtained by the discriminator based on the generated samples. If D Y The more fake the generated samples are, the more likely D Y Rating D Y The closer (G(x)) is to 0, the better the 1-D Y The closer (G(x)) is to 1, Represents the probability distribution P data(x) The average amount of information contained, Represents the probability distribution P data(y) The average amount of information contained.

[0055] For the mapping function F: X→Y and its discriminator DY , just swap x and y in the above formula to get its mapping target.

[0056] In the process of confrontation between the generator and the discriminator, the generator's ability to generate samples that are indistinguishable from the real thing becomes stronger and stronger, and eventually reaches a level that the discriminator cannot distinguish. After the adversarial training between the two, the discriminator's ability becomes stronger and stronger, and it can better distinguish the real y from G(x) generated by the generator based on x, and the loss value will be larger.

[0057] In this embodiment, a cycle consistency loss function is used for motivation, that is, x→G(x)→F(G(x))≈x is defined as forward cycle consistency, and the reverse cycle consistency is satisfied for images from domain Y: y→F(y)→G(F(y))≈y. The cycle consistency loss function is:

[0058]

[0059] Since adversarial loss alone cannot guarantee that the mapping of a single input will result in the desired output, in order to ensure the cycle consistency of the adversarial network, the above method can effectively reduce the possible mapping function space and achieve the best cycle consistency effect.

[0060] In this embodiment, the discriminator D in step S2 B When the probability of judging whether the shadow-free image B1 comes from generator A2B or the real shadow-free image B is 0.5, the training of the shadow removal network model reaches the optimal level. That is, the shadow-free image B1 generated by generator A2B is infinitely close to the real shadow-free image B, and generator A2B has reached the point where it is indistinguishable from the real thing.

[0061] In this embodiment, step S1 also includes screening of sample data, specifically: deleting sample data with incomplete images, clarity below the normal recognition range of the human eye, and unqualified content. It also includes processing the sample data, specifically: using the open CV: resize algorithm to crop or scale the image to a uniform size for all sample data, and using the Z-score standardization method to normalize the data set. The unified processing and preprocessing of the data facilitates the subsequent unified recognition of the data.

[0062] Finally, the method provided in this embodiment is applied, and the application effect is described below, as well as a comparison with the prior art.

[0063] In order to quantitatively evaluate the shadow detection performance, the provided ground mask is compared with the mask obtained by the method of the present invention. At the same time, due to the unbalanced nature of the shadow data, in order to comprehensively evaluate the performance of the method of the present invention, the following average error rate is used in the embodiment to test the effect:

[0064]

[0065] Among them, TP: True Positive (also called true positive), means that it is predicted to be a positive example and it is actually a positive example; FP: False Positive (also called false positive), means that it is predicted to be positive but it is actually a negative example; FN: False Negative (also called false negative), means that it is predicted to be a negative example but it is actually a positive example; TN: True Negative (also called true negative), means that it is predicted to be negative and it is actually a negative example; TP+FN represents all samples that are actually positive; TN+FP represents all samples that are actually negative.

[0066] After the cycle-consistent adversarial network was trained with a large number of shadow datasets, a large amount of loss data output during the training process was analyzed, and the following results were obtained: Figure 5 The adversarial loss function graph shown. Figure 5 It can be seen that the loss function has achieved relative cycle consistency, which can well ensure that the generated image retains the information of the source image as much as possible.

[0067] According to the training of adversarial network Figure 5 From the loss curve, we can see that except for the large fluctuations in the loss function values ​​of individual data during training, other losses gradually converge between 0.5 and 1 after 300 epochs of training, and the data of the loss functions output by the generator and the discriminator are basically consistent. At this point, the discriminator is gradually unable to distinguish between the generated shadow images and the shadow images in the samples, so the network can be used to generate shadow-free images.

[0068] In order to compare the effect of the present invention, the method of converting the image of RGB color space to HSV color space is adopted to eliminate shadows. In the HSV color model, the three components are independent of each other and do not affect each other. Therefore, when the V component is screened out, that is, the influence of illumination is eliminated, it does not affect the analysis of H and S components, that is, the essence of color. In practical applications, the interference of shadows can be eliminated according to this characteristic of the HSV color model. The RGB and HSV color models are very different in expression. The RGB color model is mostly used in display systems, while the HSV color model is often used in visual processing analysis. The RGB color model is converted to the HSV color model as follows:

[0069] max=max(R,G,B)

[0070] min = min(R, G, B)

[0071]

[0072]

[0073]

[0074] The shadow detection algorithm based on color features eliminates shadows according to the color difference between pixels in shadow areas and non-shadow areas. Through the color feature histogram, the number of pixels in the three components of HSV is counted and analyzed, so that the difference in pixels in the three components can be seen more intuitively. The different threshold attributes of shadows and targets in the three components are used to compare the three groups of histograms of H, S, and V with and without shadows, and the threshold is used to segment and eliminate shadows.

[0075] from Figure 6 It can be seen that when there is no shadow, the characteristics of the three components tend to be a straight line, the S component is basically unchanged, the saturation is basically maintained at 4000, and the V component has no obvious change. Only the H component is between 0.4-0.7 after the shadow appears, and the threshold above 8000 is slightly higher than that without shadow. Therefore, after converting RGB to HSV image, the method of setting the threshold cannot effectively achieve shadow removal. This is because there are complex background factors in the power training environment, which makes it difficult to process the shadows of the environment, people and shadows through accurate thresholds. The experimental results show that the personnel in the power training field are difficult to distinguish during the processing, and it is difficult to distinguish them in the processed image, so the shadow cannot be effectively eliminated, which is not conducive to subsequent analysis. Therefore, it is necessary to further study the shadow removal algorithm on this basis.

[0076] The advantage of the present invention is that it uses a cycle-consistent adversarial network to solve the shadow elimination problem. Traditional shadow elimination algorithms need to mark the shadow area of ​​the detection target during the shadow detection stage when performing shadow elimination, and cannot accurately identify the shadow area. However, the Cycle GAN adversarial network used in the present invention does not need to mark the shadow target during the game between the generator and the discriminator, and does not need to identify whether each pixel has been converted between pixels. When Cycle GAN performs shadow elimination, through a large amount of training on the shadow data set, the neural network constantly plays a self-game in the confrontation, thereby better realizing the style transfer of the shadow image, and better realizing the elimination of shadows while ensuring that the generator and the discriminator are cycle-consistent during the training process, and obtaining a good shadow elimination effect.

[0077] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A shadow elimination method in personnel image detection in an electric power training field, characterized in that: The steps include: S1. Acquire sample data, where the sample data includes a true shadow image A and a true shadow-free image B; S2. Conduct adversarial training of the generative adversarial network. The real shadow image A is used as the input data of the generator A2B. The generator A2B outputs the shadow-free image B1. The shadow-free image B1 is used as the input data of the generator B2A. The generator B2A outputs the shadow image A1. The real shadow-free image B is used as the input data of the generator B2A. The generator B2A outputs the shadow image A2. The shadow image A2 is used as the input data of the generator A2B. The generator A2B outputs the shadow-free image B2. The discriminator D A Determine whether the shadow image A2 is generated by the generator B2A or the real shadow image A, and feed back the first error between the shadow image A2 and the real shadow image A to the generator B2A, so that the generator B2A reduces the first error and improves the authenticity of the generated image; the discriminator D B Determine whether the shadow-free image B1 comes from the generator A2B or the real shadow-free image B, and feed back a second error between the shadow-free image B1 and the real shadow-free image B to the generator A2B, so that the generator A2B reduces the second error and improves the authenticity of the generated image; Generator A2B and generator B2A are respectively connected to the discriminator D B and the discriminator D A Conduct repeated adversarial training to obtain the optimal shadow removal network model; The generator A2B includes an encoder, a converter and a decoder; The encoder inputs image features extracted from the sample data by a convolutional neural network; The converter combines similar features of the image features and determines, based on the similar features, how to convert a feature vector of the image from a feature vector of a real shadow image A data set into a feature vector of a shadow-free image B data set; The decoder uses a deconvolution layer to restore low-level features from the feature vector, and converts the obtained low-level features to generate a generated image in the shadow-free image B dataset and having the same size as the real shadow image A dataset; The first error and the second error are both calculated by generating an adversarial loss of a network, and the adversarial loss is applied to a first mapping function and a second mapping function, wherein the first mapping function represents a mapping of a dataset X of shadow images to a dataset Y of shadow-free images, and the second mapping function represents a mapping of a dataset Y of a real shadow-free image B to a dataset X of a real shadow image A, and the mapping target of the first mapping function is: Among them, y represents the sample in the data set Y, x represents the sample in the data set X, and D Y (y) indicates the shadow-free image in the discriminator D Y The closer the score is to 1, the more realistic the discriminator thinks the image is. G(x) is the sample generated by the generator based on x and has the same distribution as the dataset Y. Y (G(x)) is the score obtained by the discriminator based on the generated samples. If D Y The more fake the generated samples are, the more likely D Y Rating D Y The closer (G(x)) is to 0, the better the 1-D Y The closer (G(x)) is to 1, Represents the probability distribution P data(x) The average amount of information contained, Represents the probability distribution P data(y) The average amount of information contained; A cycle consistency loss function is used for excitation, and the cycle consistency loss function is: S3, using the shadow removal network model to remove shadows, inputting images of people in the electric power training field with shadows, and outputting images of people in the electric power training field without shadows; In step S2, the discriminator D B When the probability of determining whether the shadow-free image B1 comes from the generator A2B or the real shadow-free image B is 0.5, the training of the shadow removal network model reaches the optimal level.

2. The shadow elimination method in personnel image detection in the electric power training field according to claim 1 is characterized in that: Step S1 also includes screening of sample data, specifically: deleting sample data with incomplete images, clarity below the normal recognition range of the human eye, and unqualified content.

3. The shadow elimination method in personnel image detection in an electric power training field according to claim 1 is characterized in that: Step S1 also includes processing the sample data, specifically: using the openCV: resize algorithm to crop or scale the image to a uniform size for all sample data, and using the Z-score standardization method to normalize the data set.