A power safety monitoring image detection method and system

Abstract

The application relates to the technical field of computer vision, in particular to a power safety monitoring image detection method and system, which comprises the following steps: inputting a to-be-detected power safety monitoring image and a generated image into a trained power safety monitoring image detection model respectively to obtain a detection result, the power safety monitoring image detection model comprises a feature extraction network structure, a feature distillation network structure and a multi-scale aggregation network structure in sequence, the feature extraction network structure is used for extracting local spatial features and global context information features from the to-be-detected power safety monitoring image and each generated image respectively, and fusing the two to obtain fused features corresponding to each image; the feature distillation network structure is used for extracting Value values and Key values from the fused features of each image, and obtaining splicing features based on the Value values and the Key values of each image; and the multi-scale aggregation network structure is used for processing the splicing features to obtain the detection result of the to-be-detected power safety monitoring image.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a method and system for detecting power safety monitoring images. Background Technology

[0002] With the continuous development of smart grid construction, computer vision-based image detection technology for power construction safety has become a key means to achieve intelligent monitoring of personnel behavior and real-time early warning of violations. However, in actual power construction site monitoring scenarios, this technology faces two fundamental challenges: first, specific violation samples are scarce (such as specific postures of not wearing a safety helmet, improper wearing of safety ropes, etc.), making it difficult to obtain large-scale labeled data; second, the on-site environment is complex and variable, with adverse factors such as drastic changes in lighting, severe weather interference, cluttered construction backgrounds, and dense personnel obstruction. Under the dual constraints of scarce violation samples and variable monitoring scenarios, existing intelligent safety monitoring systems struggle to meet the requirements for highly reliable early warning in terms of detection accuracy, generalization ability, and robustness. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] In view of the above-mentioned shortcomings and deficiencies of the existing technology, this application provides a power safety monitoring image detection method and system, which solves the technical problems of the extremely small number of samples of many key anomalies (such as specific equipment defects and rare violations) in the existing power safety monitoring, which cannot support the data-driven deep learning model to be fully trained, resulting in model overfitting and poor generalization ability, as well as the complex power field environment (lighting changes, occlusion, background clutter) and the easy interference of target features (such as fine cracks, safety helmets, etc.).

[0005] (II) Technical Solution

[0006] To achieve the above objectives, the main technical solutions adopted in this application include:

[0007] In a first aspect, embodiments of this application provide a method for detecting power safety monitoring images, the method comprising:

[0008] The power safety monitoring image to be detected and the generated image generated based on the pre-acquired power safety monitoring image are respectively input into the trained power safety monitoring image detection model to obtain the detection result of the power safety monitoring image to be detected. The detection result includes the category information of abnormal behavior or equipment defects in the power safety monitoring image to be detected and its corresponding location information.

[0009] The power safety monitoring image detection model sequentially includes a feature extraction network structure, a feature distillation network structure, and a multi-scale aggregation network structure, wherein:

[0010] The feature extraction network structure is used to extract local spatial features and global contextual information features from the power safety monitoring image to be detected and each generated image, and then fuse the two to obtain the fused features corresponding to each image.

[0011] The feature distillation network structure is used to extract the value and key values ​​of the fusion features of each image, and obtain the spliced ​​features based on the value and key values ​​of each image;

[0012] The multi-scale aggregation network structure is used to process the splicing features to obtain the detection result of the power safety monitoring image to be detected;

[0013] The power safety monitoring image detection model is trained by using historical power safety monitoring images and generated images based on those images.

[0014] Preferably, in some embodiments of this application, the feature extraction network structure includes a preprocessing module, a wavelet transform high-frequency fusion module, a dual-stream feature extraction module, and a feature fusion module;

[0015] The preprocessing module is used to sequentially perform low-light enhancement processing and noise reduction processing on the power safety monitoring image to be detected and the generated image to obtain the enhanced image of each image;

[0016] The wavelet transform high-frequency fusion module is used to sequentially perform discrete wavelet decomposition on the power safety monitoring image to be detected and the generated image, extract multiple high-frequency sub-band features, and perform weighted fusion of the multiple high-frequency sub-band features based on attention weights to obtain the wavelet transform feature map of each image.

[0017] The dual-stream feature extraction module includes a CNN stream for extracting local spatial features of the enhanced image and a Transformer stream for extracting global contextual information features of the wavelet transform feature map;

[0018] The CNN stream is used to perform multi-scale convolutional feature extraction on the enhanced image. The ResNet-50 network is used to extract convolutional features at different levels. Deformable convolution is introduced in the last two feature extraction stages of the ResNet-50 network to enhance the modeling ability of irregular structures. At the same time, local spatial feature maps at multiple scales are constructed based on the feature pyramid network to obtain local spatial features.

[0019] The Transformer stream is used to perform global context information feature extraction on the wavelet transform feature map, divide the wavelet transform feature map into fixed-size patches, and perform global attention modeling through a Transformer encoder with a dynamic Sigmoid gated multi-head attention mechanism to obtain global context information features.

[0020] The feature fusion module is used to unify the local spatial features and the global contextual information features to the same size at multiple scales, and perform weighted fusion based on an attention-guided fusion gating mechanism to form the fused features.

[0021] Preferably, in some embodiments of this application, the low-light enhancement processing is as follows: using the multi-scale Retinex algorithm, the illumination component is estimated, the reflection component is enhanced, and the color is restored for the power safety monitoring image to be detected and the generated image, so as to obtain the low-light enhanced image of each image respectively;

[0022] The denoising process is as follows: using a DnCNN network, residual noise prediction is performed on the low-light enhanced image of each image, and the mean square error is used as the loss function for denoising to obtain the enhanced image of each image.

[0023] Preferably, in some embodiments of this application, the feature distillation network structure includes: a value extraction unit based on channel attention mechanism, a key extraction unit based on spatial attention mechanism, a weight calculation unit, and a splicing unit;

[0024] The Value extraction unit based on the channel attention mechanism is used to extract the Value value of each image corresponding to the fusion feature in the power safety monitoring image to be detected and the generated image.

[0025] The Key value extraction unit based on the spatial attention mechanism is used to extract the Key value of each image corresponding to the fusion feature in the power safety monitoring image to be detected and the generated image.

[0026] The weight calculation unit is used to calculate the similarity between the Key value corresponding to the power safety monitoring image to be detected and the Key value corresponding to each generated image, and to obtain the weight of each generated image relative to the power safety monitoring image to be detected based on the similarity.

[0027] The stitching unit is used to multiply the Value corresponding to each generated image by the weight of the generated image relative to the power safety monitoring image to be detected, and then stitch it together with the Value of the power safety monitoring image to be detected to obtain stitching features.

[0028] Preferably, in some embodiments of this application, the multi-scale aggregation network structure includes an RPN module, a context-aware feature aggregation module, and a prediction module;

[0029] The RPN module is used to generate a set of candidate region proposals based on the splicing features;

[0030] The context-aware feature aggregation module is used to receive the candidate region proposals and their corresponding spliced ​​features, and to perform multi-scale context feature extraction and fusion on each candidate region proposal to generate a refined feature representation corresponding to each proposal.

[0031] The prediction module is used to perform classification and bounding box regression operations on each candidate region proposal based on the refined feature representation to obtain the detection result of the power safety monitoring image to be detected.

[0032] Preferably, in some embodiments of this application, the context-aware feature aggregation module includes a candidate region pooling unit, a multi-scale feature extraction unit, and an attention-weighted fusion unit;

[0033] The candidate region pooling unit is used to extract the corresponding local feature map from the spliced ​​features based on the position of each proposed candidate region.

[0034] The multi-scale feature extraction unit is used to extract features at different scales from the local feature map of each candidate region proposal to obtain the multi-scale context features of the proposal.

[0035] The attention-weighted fusion unit is used to calculate the attention weights of the multi-scale contextual features of each candidate region proposal, and perform weighted fusion accordingly to generate a refined feature representation corresponding to the proposal.

[0036] Preferably, in some embodiments of this application, the trained power safety monitoring image detection model is a power safety monitoring image detection model trained on a validation set that meets a preset performance acceptance standard.

[0037] The verification set includes historical power safety monitoring images and generated images based on historical power safety monitoring images;

[0038] The performance acceptance criterion is: the recognition accuracy of the power safety monitoring image detection model on the validation set is not lower than a preset threshold; wherein, the recognition accuracy is calculated as: the proportion of all real abnormal targets successfully detected by the power safety monitoring image detection model in the validation set, among which the targets are correctly classified and located.

[0039] Preferably, in some embodiments of this application, the generated image is generated by a generation module based on historical power safety monitoring images; wherein, the generation module employs a generative adversarial network with enhanced detail, including a generator and a discriminator:

[0040] The generator adopts a U-Net structure with skip connections, including an encoder, a decoder, a high-frequency detail recovery module, and a multi-scale feature fusion structure;

[0041] The encoder is used to downsample the input historical power safety monitoring images step by step and extract multi-scale features of each layer to form feature maps of each layer of the encoder.

[0042] The decoder is used to upsample the deepest feature map of the encoder step by step to obtain the preliminary feature map of each layer of the decoder, and to receive the skip connection features from each layer of the encoder. The upsampled preliminary feature map of each layer of the decoder is fused with the corresponding skip connection features to form a fused feature map.

[0043] The skip connection feature is the feature map extracted from each layer of the encoder;

[0044] The high-frequency detail recovery module includes at least two residual blocks, which enhance the high-frequency information of the fused feature map and output the enhanced decoder feature map.

[0045] A multi-scale feature fusion structure is used to merge the enhanced decoder feature map with the feature maps of each layer of the encoder after channel adjustment to obtain the final feature map, and then process the final feature map with a convolutional layer to generate the corresponding generated image.

[0046] The discriminator adopts the PatchGAN structure to perform local region authenticity discrimination on the generated image output by the generator, obtain the discrimination result, and through adversarial training between the generator and the discriminator, the generator can generate generated images that can be used to train the power safety monitoring image detection model.

[0047] Preferably, in some embodiments of this application, the loss function of the generator in the generation module... for:

[0048] ;

[0049] in, , , ;

[0050] ;

[0051] in, This represents the expectation of the discriminator output score for the generated image; This represents the expectation of the discriminator's output score for a real image; This indicates that the discriminator performs random interpolation. gradient at; Represents the L2 norm; Represents all samples taken from the data distribution of the random interpolation sample. Calculate the average value; It is the generated image. These are images from power safety monitoring systems. It is a random interpolation of power safety monitoring images and generated images. This is the gradient penalty coefficient, set to 10;

[0052] ;

[0053] in, This represents the feature map obtained after the input image is forward-propagated through the VGG19 network to the i-th layer. represents the weighting coefficients corresponding to the i-th layer among the five activation layers in the VGG19 network; where, , , , , They are 1 / 32, 1 / 16, 1 / 8, 1 / 4, and 1 respectively;

[0054] ;

[0055] in, This represents the Laplacian operator, used to extract high-frequency components of the input image and output a Laplacian response map; This represents the L1 norm operator.

[0056] On the other hand, embodiments of this application also provide a power safety monitoring image detection system, including:

[0057] At least one processor; and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the power safety monitoring image detection method described above.

[0058] (III) Beneficial Effects

[0059] The power safety monitoring image detection method and system provided in this application extracts the Value and Key values ​​of the fusion features of the power safety monitoring image to be detected and each generated image through a feature distillation network structure, and obtains spliced ​​features based on these values. This achieves knowledge transfer and semantic enhancement from the generated image to the power safety monitoring image to be detected, significantly improving the model's representation learning ability and generalization performance under limited sample conditions. The feature extraction network structure extracts local spatial features and global contextual information features from the power safety monitoring image to be detected and each generated image respectively, and fuses them to obtain fusion features corresponding to each image. This allows the model to simultaneously capture key local details of the target and the overall scene semantics, enhancing the discriminativeness and robustness of the features under complex backgrounds and occlusion conditions, thus providing a more reliable feature foundation for subsequent detection and effectively reducing false detections and false negatives. The spliced ​​features output by the feature distillation network structure are processed through a multi-scale aggregation network structure, which further fuses multi-scale contextual information and optimizes feature representation, ultimately obtaining the detection result of the power safety monitoring image to be detected. This improves the recognition accuracy and localization accuracy of various abnormal behaviors or equipment defects, adapting to the detection needs of power safety monitoring scenarios with varying target scales and complex environments. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of a power safety monitoring image detection method according to an embodiment of this application;

[0061] Figure 2 This is a schematic diagram of the feature distillation network structure in the power safety monitoring image detection model in one embodiment of this application. Detailed Implementation

[0062] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0063] The power safety monitoring image detection method provided in this application addresses the shortcomings of existing power safety monitoring image detection technologies in terms of sample scarcity, feature modeling, and environmental adaptability, proposing a novel detection method. Among related technologies, small-sample anomaly detection for power safety monitoring images can be mainly categorized into three types:

[0064] The first category is detection schemes based on fine-tuning pre-trained models. These schemes typically use deep neural networks pre-trained on general large-scale datasets (such as ImageNet and COCO) as the feature extraction backbone. Supervised fine-tuning is then performed on a limited number of power safety monitoring image samples to adapt to specific anomaly or defect identification tasks. However, this scheme is highly dependent on the consistency of the distribution between the pre-training data and the target domain. Since general datasets lack the structural features and anomaly patterns specific to power equipment, direct fine-tuning often makes it difficult for the model to capture key discriminative features. At the same time, when samples are extremely scarce, the model is prone to overfitting, resulting in a significant decrease in generalization ability and making it difficult to cope with complex and ever-changing detection scenarios on-site.

[0065] The second category is few-shot detection schemes based on meta-learning or metric learning. These schemes construct a support set-query set learning mechanism, enabling the model to learn to "compare" the similarity between new samples and known samples, thus quickly adapting to new categories. While such methods have certain advantages in small-shot settings, they often rely on hand-designed metric functions or simple feature embedding methods, failing to explicitly model the semantic relationships between local features within the image and the global context, and also failing to fully utilize structured prior knowledge from historical data. Therefore, in images with complex power equipment structures and strong background interference, the feature discrimination power is insufficient, resulting in poor detection stability.

[0066] The third category is data augmentation detection schemes based on generative adversarial networks (GANs). This approach uses GANs to synthesize new training samples to expand the dataset size before training a conventional detection model. While this method increases the number of samples, ordinary generative models lack specific constraints on the semantic structure and abnormal patterns of power equipment. The generated images often suffer from distorted details, semantic inconsistencies, and even the introduction of irrelevant noise, resulting in low-quality augmented data. Furthermore, data augmentation and detection models are usually trained separately, and the features of the generated images do not establish an adaptive association with the features of the real images during the detection process. This leads to low knowledge transfer efficiency and limited improvement in the final detection performance.

[0067] Therefore, the power safety monitoring image detection method provided in this application constructs a detection model by sequentially cascading a feature extraction network structure, a feature distillation network structure, and a multi-scale aggregation network structure. It also employs historical power safety monitoring images and their generated images for collaborative training. This enables high-quality sample expansion at the data level, semantic association and knowledge distillation between power safety monitoring images and generated images at the feature level, and multi-scale context-aware target localization and classification at the detection level. Thus, even in power safety monitoring scenarios where abnormal samples are scarce and environmental interference is diverse, it can still achieve high-precision and robust real-time detection of abnormal behaviors and equipment defects. The model is not only compact and end-to-end trainable, but also boasts high detection accuracy, strong generalization ability, and good environmental adaptability.

[0068] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0069] Figure 1 This is a schematic diagram of a power safety monitoring image detection method according to an embodiment of this application. Figure 1 As shown, the power safety monitoring image detection method includes:

[0070] The power safety monitoring image to be detected and the generated image generated based on the pre-acquired power safety monitoring image are respectively input into the trained power safety monitoring image detection model to obtain the detection result of the power safety monitoring image to be detected. The detection result includes the category information of abnormal behavior or equipment defects in the power safety monitoring image to be detected and its corresponding location information.

[0071] The power safety monitoring image detection model sequentially includes a feature extraction network structure, a feature distillation network structure, and a multi-scale aggregation network structure, wherein:

[0072] The feature extraction network structure is used to extract local spatial features and global contextual information features from the power safety monitoring image to be detected and each generated image, and then fuse the two to obtain the fused features corresponding to each image.

[0073] The feature distillation network structure is used to extract the value and key values ​​of the fusion features of each image, and obtain the spliced ​​features based on the value and key values ​​of each image;

[0074] The multi-scale aggregation network structure is used to process the splicing features to obtain the detection result of the power safety monitoring image to be detected;

[0075] The power safety monitoring image detection model is trained by using historical power safety monitoring images and generated images based on those images.

[0076] Specifically, the method first acquires the power safety monitoring image to be detected and inputs it, along with a generated image derived from the pre-acquired power safety monitoring image, into a pre-trained power safety monitoring image detection model. This model consists of a feature extraction network structure, a feature distillation network structure, and a multi-scale aggregation network structure, enabling collaborative processing of the power safety monitoring image to be detected and the generated image. Ultimately, it outputs the category information and precise location information of the target in the power safety monitoring image to be detected. During the training phase, the model uses historical power safety monitoring images and their corresponding generated images for end-to-end training, thereby learning the ability to extract robust features from limited samples and perform reliable inference. Firstly, at the data level, this embodiment effectively alleviates the problem of scarce abnormal samples in power safety monitoring scenarios by introducing generated images as training samples. At the feature fusion level, the feature distillation network structure extracts and associates the key and value values ​​of the power safety monitoring image to be detected and the generated image, achieving cross-sample semantic knowledge distillation and transfer. This mechanism enables the model to extract useful prior features from generated images and adaptively fuse them with the features of the power safety monitoring images to be detected, thereby enhancing the model's representation learning ability and generalization performance under small sample conditions and effectively suppressing overfitting. Secondly, at the feature representation level, the feature extraction network structure adopts a dual-stream mechanism, extracting local spatial features and global contextual information features in parallel and deeply fusing them. Local spatial features focus on the detailed texture and shape information of targets such as safety helmets, insulator cracks, and safety rope buckles, while global contextual information features capture the overall semantic relationship between equipment layout, personnel, and the environment. The fusion of these two features ensures that the features generated by the model possess both detail discrimination and scene understanding capabilities, thus maintaining high discriminative power even when facing complex backgrounds, occlusion interference, and multi-scale targets, significantly reducing false positives and false negatives. Furthermore, at the detection optimization level, a multi-scale aggregation network structure performs deep processing on the spliced ​​features output by the feature distillation network structure. Based on the established semantic relationships between features, this structure performs context-aware multi-scale feature extraction and fusion on candidate target regions, achieving further optimization and refinement of the initial features. This process effectively filters out background noise, enhances the multi-scale representation of the target, and enables the model to adapt to the target detection requirements of different sizes and different occlusion states, thereby improving the detection robustness and positioning accuracy in variable field environments.

[0077] In practical applications of this application, the feature extraction network structure includes a preprocessing module, a wavelet transform high-frequency fusion module, a dual-stream feature extraction module, and a feature fusion module;

[0078] The preprocessing module is used to sequentially perform low-light enhancement processing and noise reduction processing on the power safety monitoring image to be detected and the generated image to obtain the enhanced image of each image;

[0079] In this embodiment, the low-light enhancement processing is as follows: using the multi-scale Retinex algorithm, the illumination component is estimated, the reflection component is enhanced, and the color is restored for the power safety monitoring image to be detected and the generated image, so as to obtain the low-light enhanced image of each image respectively;

[0080] The denoising process is as follows: using a DnCNN network, residual noise prediction is performed on the low-light enhanced image of each image, and the mean square error is used as the loss function for denoising to obtain the enhanced image of each image.

[0081] The wavelet transform high-frequency fusion module is used to sequentially perform discrete wavelet decomposition on the power safety monitoring image to be detected and the generated image, extract multiple high-frequency sub-band features, and perform weighted fusion of the multiple high-frequency sub-band features based on attention weights to obtain the wavelet transform feature map of each image.

[0082] The dual-stream feature extraction module includes a CNN stream for extracting local spatial features of the enhanced image and a Transformer stream for extracting global contextual information features of the wavelet transform feature map;

[0083] The CNN stream is used to perform multi-scale convolutional feature extraction on the enhanced image. The ResNet-50 network is used to extract convolutional features at different levels. Deformable convolution is introduced in the last two feature extraction stages of the ResNet-50 network to enhance the modeling ability of irregular structures. At the same time, local spatial feature maps at multiple scales are constructed based on the feature pyramid network to obtain local spatial features.

[0084] The Transformer stream is used to perform global context information feature extraction on the wavelet transform feature map, divide the wavelet transform feature map into fixed-size patches, and perform global attention modeling through a Transformer encoder with a dynamic Sigmoid gated multi-head attention mechanism to obtain global context information features.

[0085] The feature fusion module is used to unify the local spatial features and the global contextual information features to the same size at multiple scales, and perform weighted fusion based on an attention-guided fusion gating mechanism to form the fused features.

[0086] Specifically, the feature fusion module is responsible for deeply fusing the local spatial features output from the CNN stream with the global contextual information features output from the Transformer stream. This module first unifies the two types of features to the same size across multiple scales. Then, through an attention-guided fusion gating mechanism, it dynamically calculates the fusion weights for each location and channel, achieving adaptive weighted integration of local details and global semantics. The resulting fused features contain both local representations of fine structures such as insulator cracks and safety helmet buckles, and high-level semantic information such as equipment layout and personnel behavior relationships, thus enhancing the environmental adaptability and semantic consistency of the features while preserving discriminative details.

[0087] Preferably, in some embodiments of this application, see Figure 2 The feature distillation network structure includes: a value extraction unit based on channel attention mechanism, a key extraction unit based on spatial attention mechanism, a weight calculation unit, and a splicing unit;

[0088] The Value Extraction Unit, based on a channel attention mechanism, extracts the Value from the fused features of each image in the power safety monitoring image to be detected and the generated image. Here, the Value can be understood as the information carrier of actual semantic content within each image feature; for example, it may encode visual content such as a specific texture pattern of a type of safety helmet or the morphological details of an insulator crack. By applying attention weights to the channel dimensions, this unit strengthens the feature channels important to the recognition task and suppresses irrelevant or redundant information, thereby obtaining a more discriminative set of content features.

[0089] The Key value extraction unit, based on a spatial attention mechanism, extracts key values ​​from the fused features corresponding to each image in the power safety monitoring image to be detected and the generated image. The key value acts like an "index" or "identifier," reflecting the semantic attributes or contextual relationships of features at different spatial locations. For example, it might characterize a region as belonging to a "high-altitude work area," "near a grounding device," or "densely populated area." Through the spatial attention mechanism, the model can focus on those discriminative key regions, enabling the extracted key values ​​to more accurately represent the semantic layout of the image.

[0090] The weight calculation unit is used to calculate the similarity between the Key value corresponding to the power safety monitoring image to be detected and the Key value corresponding to each generated image, and to obtain the weight of each generated image relative to the power safety monitoring image to be detected based on the similarity.

[0091] The weight calculation unit calculates the similarity between the Key value corresponding to the power safety monitoring image to be detected and the Key value corresponding to each generated image. This similarity measures the degree of matching between the generated image and the power safety monitoring image to be detected in terms of semantic layout and contextual relationships. For example, if a generated image depicts a similar work scenario to the power safety monitoring image to be detected (such as both including an aerial platform and personnel interaction), its Key value similarity is high; conversely, if the scenario is significantly different, the similarity is low. Based on this similarity, the unit assigns a weight relative to the power safety monitoring image to be detected to each generated image. A higher weight indicates a stronger semantic association between the generated image and the power safety monitoring image to be detected, and the more transferable the knowledge it carries is.

[0092] The stitching unit is used to multiply the Value corresponding to each generated image by the weight of the generated image relative to the power safety monitoring image to be detected, and then stitch it together with the Value of the power safety monitoring image to be detected to obtain stitching features.

[0093] Specifically, the stitching unit multiplies the Value corresponding to each generated image with the weight acquired by that image, achieving weighted filtering and modulation of knowledge. Then, these weighted Values ​​are stitched together with the Value of the power safety monitoring image to be detected, forming the final stitched features. This process essentially involves selectively fusing semantically relevant and useful visual knowledge (carried by Values) from multiple generated images that is relevant to the power safety monitoring image to be detected, and supplementing it into the original features. For example, if a generated image provides details about the tilt angle of a safety helmet that is missing in the power safety monitoring image to be detected, and this image receives a high weight due to scene similarity, then its detailed features will be enhanced and fused into the stitched features, thereby enhancing the model's ability to recognize similar situations.

[0094] In this embodiment, selective knowledge transfer is achieved through a feature distillation network structure. The weight calculation unit ensures that only generated image knowledge highly relevant to the semantics of the power safety monitoring image to be detected is fused, avoiding the introduction of irrelevant or conflicting information. Furthermore, it supports model generalization under small sample conditions. Even with few anomalous samples, the model can learn more robust feature representations by associating relevant patterns in a large number of generated images, thus maintaining high detection accuracy when facing new scenarios and situations.

[0095] In this embodiment, the multi-scale aggregation network structure includes an RPN module, a context-aware feature aggregation module, and a prediction module;

[0096] The RPN module is used to generate a set of candidate region proposals based on the stitching features. For example, the set of candidate region proposals generated by the RPN module is a series of rectangular boxes with preliminary location and size information. For instance, in a power construction scene image, the RPN may simultaneously propose the area of ​​a worker's head in the distance (small target), the area of ​​a toolbox in the foreground (medium target), and the area of ​​large equipment (large target), ensuring that potential targets of various scales are not missed.

[0097] The context-aware feature aggregation module is used to receive the candidate region proposals and their corresponding spliced ​​features, and to perform multi-scale context feature extraction and fusion on each candidate region proposal to generate a refined feature representation corresponding to each proposal.

[0098] Specifically, the context-aware feature aggregation module includes a candidate region pooling unit, a multi-scale feature extraction unit, and an attention-weighted fusion unit;

[0099] The candidate region pooling unit is used to extract the corresponding local feature map from the splicing features based on the location of each proposed candidate region; for example, for a proposed "helmet" region, the unit will extract the local feature map of the region and a certain range around it.

[0100] The multi-scale feature extraction unit is used to extract features at different scales from the local feature map of each candidate region proposal to obtain the multi-scale context features of the proposal; for example, by using dilated convolution with different dilation rates or pooling operations of different scales, a feature set that characterizes the context information of different ranges of the local feature map can be obtained—which helps to understand the details of the "safety helmet" itself (small-scale features) as well as its relationship with the surrounding people and equipment (large-scale features).

[0101] The attention-weighted fusion unit is used to propose and calculate the attention weights of the multi-scale contextual features for each candidate region, and then perform weighted fusion to generate a refined feature representation corresponding to the proposal. For example, for a small-sized "insulator crack," the model may focus more on small-to-medium-scale details; while for a partially occluded "safety rope," large-scale contextual features (such as connection point information) may be more critical. Through weighted fusion, the module generates a refined feature representation that integrates the most relevant multi-scale contextual information, significantly enhancing the discriminative power and robustness of the features.

[0102] The prediction module performs classification and bounding box regression operations on each candidate region proposal based on the refined feature representation to obtain the detection result of the power safety monitoring image to be detected. Specifically, the prediction module receives the refined feature representation corresponding to each candidate region proposal and performs two tasks in parallel: one is a classification operation to determine what kind of abnormal behavior or equipment defect the region belongs to (e.g., "not wearing a safety helmet," "damaged insulator"), and the other is a bounding box regression operation to fine-tune the position and size of the proposed region's bounding box to better match the boundary of the actual target. After all processing is complete, the module outputs the final detection result of the power safety monitoring image to be detected, i.e., the category and precise location of each identified target.

[0103] Among them, the trained power safety monitoring image detection model is a power safety monitoring image detection model trained on the validation set that meets the preset performance acceptance standards.

[0104] The verification set includes historical power safety monitoring images and generated images based on historical power safety monitoring images;

[0105] The performance acceptance criterion is: the recognition accuracy of the power safety monitoring image detection model on the validation set is not lower than a preset threshold; wherein, the recognition accuracy is calculated as: the proportion of all real abnormal targets successfully detected by the power safety monitoring image detection model in the validation set, among which the targets are correctly classified and located.

[0106] Specifically, the trained power safety monitoring image detection model refers to a model that is evaluated on a validation set independent of the training data and meets a preset performance acceptance standard. The validation set includes historical power safety monitoring images and generated images based on those historical images. This ensures that the validation data reflects both the distribution of real-world scenes and the semantic diversity introduced by the generated images, thus comprehensively testing the model's generalization ability. The quantitative indicator of the performance acceptance standard is the model's recognition accuracy on the validation set, which must be no less than a pre-set threshold (e.g., 95%). The calculation method for this recognition accuracy is rigorous and closely aligned with actual detection needs: the numerator is the number of all real abnormal targets in the validation set that were successfully detected by the model and whose classification and localization are both correct; the denominator is the total number of all real abnormal targets in the validation set. "Correct classification" means that the category predicted by the model (e.g., "not wearing a safety helmet," "insulator flashover trace") is consistent with the manually labeled real category; "correct localization" typically requires that the intersection-union ratio (IU) between the model's predicted bounding box and the real labeled box be no less than a certain strict threshold (e.g., 0.5). This calculation method simultaneously assesses the model's ability to detect targets, identify categories, and pinpoint locations, avoiding potential biases that might arise from evaluating the model solely based on classification accuracy or the number of detections.

[0107] Preferably, in one embodiment of this application, the generated image is generated by a generation module based on historical power safety monitoring images; wherein, the generation module employs a generative adversarial network with enhanced detail, comprising a generator and a discriminator:

[0108] The generator adopts a U-Net structure with skip connections, including an encoder, a decoder, a high-frequency detail recovery module, and a multi-scale feature fusion structure;

[0109] The encoder is used to downsample the input historical power safety monitoring images step by step and extract multi-scale features of each layer to form feature maps of each layer of the encoder.

[0110] The decoder is used to upsample the deepest feature map of the encoder step by step to obtain the preliminary feature map of each layer of the decoder, and to receive the skip connection features from each layer of the encoder. The upsampled preliminary feature map of each layer of the decoder is fused with the corresponding skip connection features to form a fused feature map.

[0111] The skip connection feature is the feature map extracted from each layer of the encoder;

[0112] The high-frequency detail recovery module includes at least two residual blocks, which enhance the high-frequency information of the fused feature map and output the enhanced decoder feature map.

[0113] A multi-scale feature fusion structure is used to merge the enhanced decoder feature map with the feature maps of each layer of the encoder after channel adjustment to obtain the final feature map, and then process the final feature map with a convolutional layer to generate the corresponding generated image.

[0114] The discriminator adopts the PatchGAN structure to perform local region authenticity discrimination on the generated image output by the generator, obtain the discrimination result, and through adversarial training between the generator and the discriminator, the generator can generate generated images that can be used to train the power safety monitoring image detection model.

[0115] Specifically, the generation module employs a Depth-Enhanced Generative Adversarial Network (DESRGAN), which includes a generator and a discriminator.

[0116] The generator adopts a U-Net structure with skip connections, including an encoder, decoder, high-frequency detail recovery module, and multi-scale feature fusion structure, as detailed below:

[0117] The encoder is used to perform step-by-step downsampling and feature abstraction on the input power safety monitoring image. The encoder consists of 5 downsampling blocks, with an input image size of 512×512×3. The specific structure is as follows: Convolutional Layer 1 (64 4×4 filters, stride 2, padding 1, followed by batch normalization and LeakyReLU), Convolutional Layer 2 (128 4×4 filters, stride 2, padding 1, batch normalization, LeakyReLU), Convolutional Layer 3 (256 4×4 filters, stride 2, padding 1, batch normalization, LeakyReLU), Convolutional Layer 4 (512 4×4 filters, stride 2, padding 1, batch normalization, LeakyReLU), and Convolutional Layer 5 (512 4×4 filters, stride 2, padding 1, batch normalization, LeakyReLU). The output of each layer constitutes the feature map of that layer in the encoder, used for subsequent skip connections and multi-scale fusion.

[0118] Decoder and Skip Connections: The decoder is used to progressively upsample the deepest feature maps of the encoder. It consists of 5 upsampling blocks, each with a core deconvolution operation (e.g., deconvolution layer 1: 512 4×4 filters, stride 2, padding 1, followed by batch normalization and ReLU). Skip connections concatenate the initial feature maps from each upsampled layer of the decoder with the corresponding feature maps from the encoder layer, forming a fused feature map.

[0119] High-Frequency Detail Restoration Module: This module consists of two residual blocks that act on the fused feature map generated by each layer of the decoder, aiming to enhance high-frequency details. Each residual block contains two convolutional layers: a first convolutional layer (64 3×3 filters, stride 1, padding 1, followed by ReLU activation) and a second convolutional layer (64 3×3 filters, stride 1, padding 1). The module's input is the fused feature map of the current layer of the decoder, and its output is added to the original fused feature map through residual connections to obtain the enhanced decoder feature map.

[0120] Multi-scale feature fusion structure: To fully utilize the multi-scale information captured during the encoding process, this structure extracts feature maps from layers 2, 3, 4, and 5 of the encoder. After adjusting the number of channels using a 1×1 convolutional layer, these maps are concatenated and fused with the enhanced decoder feature maps from the corresponding layers of the decoder. This operation achieves cross-scale feature reuse and information complementarity, generating a final feature map rich in multi-granular semantics. This final feature map is then processed by a convolutional layer (using the Tanh activation function), outputting a generated image with a size of 512×512×3.

[0121] The discriminator employs a PatchGAN architecture to determine the local realism of generated images. It consists of five stacked convolutional layers: Convolutional Layer 1 (64 4×4 layers, stride 2, padding 1, LeakyReLU), Convolutional Layer 2 (128 4×4 layers, stride 2, padding 1, batch normalization, LeakyReLU), Convolutional Layer 3 (256 4×4 layers, stride 2, padding 1, batch normalization, LeakyReLU), Convolutional Layer 4 (512 4×4 layers, stride 2, padding 1, batch normalization, LeakyReLU), and Convolutional Layer 5 (1 4×4 layer, stride 1, padding 1, Sigmoid). The discriminator outputs a two-dimensional matrix, where each element represents the probability that a corresponding local region in the image is a real image.

[0122] The Adam optimizer was used for training, with the learning rate for both the generator and discriminator set to 0.0001, a batch size of 8, and a total of 100,000 training iterations. During training, an alternating update strategy was employed: after every 5 parameter updates for the generator, the discriminator was updated once to maintain dynamic equilibrium during training.

[0123] In this embodiment, through adversarial training between the generator and the discriminator, the generator continuously learns to generate more realistic and detailed images to "deceive" the discriminator, while the discriminator continuously improves its ability to distinguish between real and fake images. This process can efficiently generate high-quality, semantically consistent augmented samples. Compared to simple geometric transformations or color jittering, this generation module can synthesize new images with reasonable physical appearance, correct contextual relationships (e.g., safety helmets always appear in the head area of ​​a person), and rich details, greatly improving the effectiveness and diversity of the augmented data. In addition, it particularly enhances the detailed information that is crucial to the detection task. Through a dedicated high-frequency detail recovery module and a locally sensitive PatchGAN discriminator, the generated images can retain or even enhance those subtle features that are crucial for distinguishing between "normal" and "abnormal" (e.g., the state of the safety helmet buckle, the morphology of insulator cracks), making the detection model trained on this data more sensitive to details.

[0124] In this embodiment of the application, the loss function L of the generator in the generation module G for:

[0125] ;

[0126] in, , , ;

[0127] ;

[0128] in, This represents the expectation of the discriminator output score for the generated image; This represents the expectation of the discriminator's output score for a real image; This indicates that the discriminator performs random interpolation. gradient at; Represents the L2 norm; Represents all samples taken from the data distribution of the random interpolation sample. Calculate the average value; It is the generated image. These are images from power safety monitoring systems. It is a random interpolation of power safety monitoring images and generated images. This is the gradient penalty coefficient, set to 10; in this embodiment of the application, The gradient penalty term forces the discriminator to perform random interpolation between the real and generated images. The gradient norm is close to 1 to stabilize the adversarial training process and prevent mode collapse and gradient vanishing. For example, when generating images simulating "not wearing a helmet," this loss causes the generator to synthesize a human head region that is consistent with the overall style of the real image in terms of lighting, texture, and integration with the environment, making it difficult for the discriminator to distinguish.

[0129] ;

[0130] in, This represents the feature map obtained after the input image is forward-propagated through the VGG19 network to the i-th layer. represents the weighting coefficients corresponding to the i-th layer among the five activation layers in the VGG19 network; where, , , , , The weights are 1 / 32, 1 / 16, 1 / 8, 1 / 4, and 1, respectively. In this embodiment, a pre-trained VGG19 network is used to constrain the deep semantic features of the generated image to be close to those of the power safety monitoring image at multiple feature layers. Specifically, the calculation is a weighted sum of the differences between the feature maps of each layer, where the weights are... As the network depth increases (1 / 32, 1 / 16, 1 / 8, 1 / 4, 1), this means that the consistency requirements for high-level semantic features become more stringent. Taking the generation of an image with simulated "insulator cracks" as an example, this loss ensures that the generated image is consistent with the power safety monitoring image in terms of scene layout, equipment type, and overall structure (high-level semantics), while allowing reasonable variations in local textures (low-level features), thereby ensuring that the generated content is semantically faithful to the original scene and avoiding illogical objects.

[0131] ;

[0132] in, This represents the Laplacian operator, used to extract high-frequency components of the input image and output a Laplacian response map; This represents the L1 norm operator. This loss function directly optimizes the fidelity of local image details. In power safety monitoring scenarios, many key anomalies (such as small cracks, wear marks, and the status of safety helmet buckles) rely on high-frequency information for identification. For example, when generating an image of "metal fitting corrosion," this loss function forces the generator not only to simulate rust in color but also to match the high-frequency features of real rust in details such as edge sharpness and texture roughness. This ensures that the generated defect area has sufficient visual discriminative power and can be effectively used to train subsequent detection models.

[0133] The generator's loss function L G The co-guided generator produces augmented images that are both realistic and rich in discriminative details. These images can provide high-quality, highly relevant training data for subsequent power safety monitoring image detection models, directly helping to improve the feature learning ability, generalization performance, and final recognition accuracy of the detection model under conditions of scarce samples.

[0134] This application also provides a power safety monitoring image detection system, including: at least one processor; and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the power safety monitoring image detection method as described in the above embodiments by calling the program instructions.

[0135] This application discloses a power safety monitoring image detection method and system. By introducing generated images during the training phase, it effectively expands the samples available for model training, alleviating the problem of insufficient model training caused by the scarcity of samples of abnormal behavior or equipment defects. Furthermore, the feature distillation network structure extracts the Value and Key values ​​of the fusion features from the power safety monitoring image and each generated image, and obtains spliced ​​features based on these values. This achieves knowledge transfer and semantic enhancement from the generated image to the power safety monitoring image, significantly improving the model's representation learning ability and generalization performance under limited sample conditions. In addition, the feature extraction network structure extracts local spatial features and global contextual information features from the power safety monitoring image and each generated image respectively, and fuses them to obtain fusion features corresponding to each image. This allows the model to simultaneously capture key local details of the target and the overall scene semantics, enhancing the discriminativeness and robustness of the features under complex backgrounds and occlusion conditions, thus providing a more reliable feature foundation for subsequent detection and effectively reducing false positives and false negatives. Finally, by processing the spliced ​​features output by the feature distillation network structure through a multi-scale aggregation network structure, multi-scale contextual information can be further integrated to optimize feature representation, thereby finally obtaining the detection results of the power safety monitoring image to be detected. This improves the recognition accuracy and positioning accuracy of various abnormal behaviors or equipment defects, and adapts to the detection needs of power safety monitoring scenarios with variable target scales and complex environments.

[0136] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0137] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0138] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0139] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A power safety monitoring image detection method, characterized in that, The method includes: The power safety monitoring image to be detected and the generated image generated based on the pre-acquired power safety monitoring image are respectively input into the trained power safety monitoring image detection model to obtain the detection result of the power safety monitoring image to be detected. The detection result includes the category information of abnormal behavior or equipment defects in the power safety monitoring image to be detected and its corresponding location information. The power safety monitoring image detection model sequentially includes a feature extraction network structure, a feature distillation network structure, and a multi-scale aggregation network structure, wherein: The feature extraction network structure is used to extract local spatial features and global contextual information features from the power safety monitoring image to be detected and each generated image, and then fuse the two to obtain the fused features corresponding to each image. The feature distillation network structure is used to extract the value and key values ​​of the fusion features of each image, and obtain the spliced ​​features based on the value and key values ​​of each image; The multi-scale aggregation network structure is used to process the splicing features to obtain the detection result of the power safety monitoring image to be detected; Among them, the power safety monitoring image detection model is trained in advance using historical power safety monitoring images and generated images based on historical power safety monitoring images to obtain the trained power safety monitoring image detection model; The feature distillation network structure includes: a value extraction unit based on channel attention mechanism, a key extraction unit based on spatial attention mechanism, a weight calculation unit, and a splicing unit; The Value extraction unit based on the channel attention mechanism is used to extract the Value value of each image corresponding to the fusion feature in the power safety monitoring image to be detected and the generated image. The Key value extraction unit based on the spatial attention mechanism is used to extract the Key value of each image corresponding to the fusion feature in the power safety monitoring image to be detected and the generated image. The weight calculation unit is used to calculate the similarity between the Key value corresponding to the power safety monitoring image to be detected and the Key value corresponding to each generated image, and to obtain the weight of each generated image relative to the power safety monitoring image to be detected based on the similarity. The stitching unit is used to multiply the Value corresponding to each generated image by the weight of the generated image relative to the power safety monitoring image to be detected, and then stitch it together with the Value of the power safety monitoring image to be detected to obtain stitching features.

2. The power safety monitoring image detection method according to claim 1, characterized in that, The feature extraction network structure includes a preprocessing module, a wavelet transform high-frequency fusion module, a dual-stream feature extraction module, and a feature fusion module. The preprocessing module is used to sequentially perform low-light enhancement processing and noise reduction processing on the power safety monitoring image to be detected and the generated image to obtain the enhanced image of each image; The wavelet transform high-frequency fusion module is used to sequentially perform discrete wavelet decomposition on the power safety monitoring image to be detected and the generated image, extract multiple high-frequency sub-band features, and perform weighted fusion of the multiple high-frequency sub-band features based on attention weights to obtain the wavelet transform feature map of each image. The dual-stream feature extraction module includes a CNN stream for extracting local spatial features of the enhanced image and a Transformer stream for extracting global contextual information features of the wavelet transform feature map; The CNN stream is used to perform multi-scale convolutional feature extraction on the enhanced image. The ResNet-50 network is used to extract convolutional features at different levels. Deformable convolution is introduced in the last two feature extraction stages of the ResNet-50 network to enhance the modeling ability of irregular structures. At the same time, local spatial feature maps at multiple scales are constructed based on the feature pyramid network to obtain local spatial features. The Transformer stream is used to perform global context information feature extraction on the wavelet transform feature map, divide the wavelet transform feature map into fixed-size patches, and perform global attention modeling through a Transformer encoder with a dynamic Sigmoid gated multi-head attention mechanism to obtain global context information features. The feature fusion module is used to unify the local spatial features and the global contextual information features to the same size at multiple scales, and perform weighted fusion based on an attention-guided fusion gating mechanism to form the fused features.

3. The power safety monitoring image detection method according to claim 2, characterized in that, The low-light enhancement process is as follows: using the multi-scale Retinex algorithm, the illumination component is estimated, the reflection component is enhanced, and the color is restored for the power safety monitoring image to be detected and the generated image, so as to obtain the low-light enhanced image of each image. The denoising process is as follows: using a DnCNN network, residual noise prediction is performed on the low-light enhanced image of each image, and the mean square error is used as the loss function for denoising to obtain the enhanced image of each image.

4. The power safety monitoring image detection method according to claim 3, characterized in that, The multi-scale aggregation network structure includes an RPN module, a context-aware feature aggregation module, and a prediction module; The RPN module is used to generate a set of candidate region proposals based on the splicing features; The context-aware feature aggregation module is used to receive the candidate region proposals and their corresponding spliced ​​features, and to perform multi-scale context feature extraction and fusion on each candidate region proposal to generate a refined feature representation corresponding to each proposal. The prediction module is used to perform classification and bounding box regression operations on each candidate region proposal based on the refined feature representation to obtain the detection result of the power safety monitoring image to be detected.

5. The power safety monitoring image detection method according to claim 4, characterized in that, The context-aware feature aggregation module includes a candidate region pooling unit, a multi-scale feature extraction unit, and an attention-weighted fusion unit. The candidate region pooling unit is used to extract the corresponding local feature map from the spliced ​​features based on the position of each proposed candidate region. The multi-scale feature extraction unit is used to extract features at different scales from the local feature map of each candidate region proposal to obtain the multi-scale context features of the proposal. The attention-weighted fusion unit is used to calculate the attention weights of the multi-scale contextual features of each candidate region proposal, and perform weighted fusion accordingly to generate a refined feature representation corresponding to the proposal.

6. The power safety monitoring image detection method according to claim 5, characterized in that, in, The trained power safety monitoring image detection model is a power safety monitoring image detection model trained on the validation set that meets the preset performance acceptance standards. The verification set includes historical power safety monitoring images and generated images based on historical power safety monitoring images; The performance acceptance criterion is: the recognition accuracy of the power safety monitoring image detection model on the validation set is not lower than a preset threshold; wherein, the recognition accuracy is calculated as: the proportion of all real abnormal targets successfully detected by the power safety monitoring image detection model in the validation set, among which the targets are correctly classified and located.

7. The power safety monitoring image detection method according to claim 6, characterized in that, The generated image is generated by the generation module based on historical power safety monitoring images; wherein, the generation module employs a detail-enhanced generative adversarial network, including a generator and a discriminator: The generator adopts a U-Net structure with skip connections, including an encoder, a decoder, a high-frequency detail recovery module, and a multi-scale feature fusion structure; The encoder is used to downsample the input historical power safety monitoring images step by step and extract multi-scale features of each layer to form feature maps of each layer of the encoder. The decoder is used to upsample the deepest feature map of the encoder step by step to obtain the preliminary feature map of each layer of the decoder, and to receive the skip connection features from each layer of the encoder. The upsampled preliminary feature map of each layer of the decoder is fused with the corresponding skip connection features to form a fused feature map. The skip connection feature is the feature map extracted from each layer of the encoder; The high-frequency detail recovery module includes at least two residual blocks, which enhance the high-frequency information of the fused feature map and output the enhanced decoder feature map. A multi-scale feature fusion structure is used to merge the enhanced decoder feature map with the feature maps of each layer of the encoder after channel adjustment to obtain the final feature map, and then process the final feature map with a convolutional layer to generate the corresponding generated image. The discriminator adopts the PatchGAN structure to perform local region authenticity discrimination on the generated image output by the generator, obtain the discrimination result, and through adversarial training between the generator and the discriminator, the generator can generate generated images that can be used to train the power safety monitoring image detection model.

8. The power safety monitoring image detection method according to claim 7, characterized in that, The loss function of the generator in the generation module for: ； in, , , ; ； in, This represents the expectation of the discriminator output score for the generated image; This represents the expectation of the discriminator's output score for a real image; This indicates that the discriminator performs random interpolation. gradient at; Represents the L2 norm; Represents all samples taken from the data distribution of the random interpolation sample. Calculate the average value; It is the generated image. These are images from power safety monitoring systems. It is a random interpolation of power safety monitoring images and generated images. This is the gradient penalty coefficient, set to 10; ； in, This represents the feature map obtained after the input image is forward-propagated to the i-th layer of the VGG19 network. represents the weighting coefficients corresponding to the i-th layer among the five activation layers in the VGG19 network; where, , , , , They are 1 / 32, 1 / 16, 1 / 8, 1 / 4, and 1 respectively; ； in, This represents the Laplacian operator, used to extract high-frequency components of the input image and output a Laplacian response map; This represents the L1 norm operator.

9. A power safety monitoring image detection system, characterized in that, include: At least one processor; as well as At least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the power safety monitoring image detection method as described in any one of claims 1-8.

Images