Power distribution room equipment image detection method, system, equipment and readable storage medium

Abstract

The invention relates to a power distribution room equipment image detection method and system, equipment and a readable storage medium. The method comprises the steps of obtaining a to-be-detected image of power distribution room equipment; inputting the to-be-detected image into a pre-trained image enhancement model, and performing enhancement processing on the target structure feature in the to-be-detected image to obtain an enhanced image; wherein the image enhancement model is a pre-trained model based on the imaging environment parameters of the distribution room; and inputting the enhanced image into the target detection model, and identifying and outputting a target region corresponding to the target structure feature. By adopting the method, the problems of low equipment detection accuracy and inaccurate positioning in a low-illumination and strong-noise environment can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment management, in particular to a power distribution room equipment image detection method and system, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] With the improvement of the intelligentization and automation level of the power system, the real-time monitoring of the equipment operation state of the power distribution room, as the core hub of power transmission and distribution, has become the key to ensuring the stable operation of the power system. In recent years, computer vision technology has been widely used in industrial detection. The image-based equipment detection method gradually replaces traditional manual inspection and becomes an important technical means for power distribution room equipment state monitoring due to its non-contact, high efficiency, wide coverage and other advantages.

[0003] In the traditional technology, the power distribution room equipment image detection mainly adopts the mode of directly detecting the original image or detecting after simple image processing: some schemes directly input the original image into the target detection model without optimizing the image quality; a few schemes only adjust the brightness or suppress noise through traditional image enhancement algorithms such as gray stretching and histogram equalization, and then perform detection.

[0004] However, the current power distribution room equipment image detection has significant limitations: the lighting conditions in the power distribution room are complex, and problems such as low light, backlight, and reflection are prone to occur, resulting in blurred image details and insufficient contrast, making it difficult for traditional enhancement algorithms to strengthen the equipment target structure features; secondly, the types of power distribution room equipment are diverse and densely arranged, and in complex working environments, the quality of image signals is easily affected by noise, resulting in insufficient anti-interference ability, high false detection rate and missed detection rate in actual engineering applications, and it is difficult to meet the engineering needs of intelligent monitoring of the power distribution room. SUMMARY

[0005] Therefore, it is necessary to provide an image detection method, device, computer equipment, computer readable storage medium and computer program product capable of accurately detecting power distribution cabinet equipment in a low-light and strong-noise environment.

[0006] In a first aspect, the present application provides a power distribution room equipment image detection method, comprising:

[0007] obtaining a to-be-detected image of the power distribution room equipment;

[0008] inputting the to-be-detected image into a pre-trained image enhancement model to enhance the target structure features in the to-be-detected image and obtain an enhanced image;

[0009] The image enhancement model is a model pre-trained based on the imaging environment parameters of the power distribution room.

[0010] inputting the enhanced image into a target detection model, identifying and outputting a target region corresponding to the target structure feature.

[0011] In one of the embodiments, the image enhancement model is trained by using a training data set, the training data set includes training images obtained by performing interference processing based on the power distribution room imaging environment parameters, and the training images are generated in the following manner:

[0012] According to the power distribution room imaging environment parameters, set the brightness attenuation threshold and the environmental medium concentration range;

[0013] Obtain the original image of the power distribution room equipment, perform random geometric transformation on the original image, and superimpose a low-illumination feature and a smoke noise layer on the original image according to the brightness attenuation threshold and the environmental medium concentration range, to generate the training image.

[0014] In one of the embodiments, the image enhancement model adopts a low-light image enhancement network fused with a channel attention mechanism;

[0015] The enhancement processing of the target structure feature in the to-be-detected image includes:

[0016] The low-light image enhancement network is used to decompose the to-be-detected image into an illumination component and a reflection component;

[0017] The channel attention mechanism is embedded into a backbone network for processing the reflection component, and a feature map of the reflection component is extracted through a convolution layer of the backbone network;

[0018] The feature map contains multiple channels, the channel presenting local high frequency and high contrast is determined as a target channel where the target structure feature is located, the weight of the target channel is increased, and a weighted reflection component is formed;

[0019] The weighted reflection component is combined with the illumination component to reconstruct and generate the enhanced image.

[0020] In one of the embodiments, the channel attention mechanism adopts an efficient channel attention network module:

[0021] The channel attention mechanism is embedded into a backbone network for processing the reflection component, and a feature map of the reflection component is extracted through a convolution layer of the backbone network, including:

[0022] An information interaction between adjacent channels in the feature map is calculated through a one-dimensional convolution kernel of the efficient channel attention network module to obtain a correlation channel feature vector;

[0023] performing compression function activation on the associated channel feature vector to generate a channel weight vector;

[0024] performing multiplication between the channel weight vector and the feature map in the channel dimension to obtain a weighted reflection component.

[0025] In one of the embodiments, the target structure feature is a local high-frequency, high-contrast, and specific-direction texture feature in the feature map of the reflection component.

[0026] The channel attention mechanism increases the weight of the target channel where the target structure feature is located, and reduces the weight of the channel corresponding to the false texture introduced by low illumination and noise.

[0027] In one of the embodiments, the target detection model is an end-to-end target detection model trained by labeled data.

[0028] The inputting of the enhanced image into the target detection model to identify and output the target region corresponding to the target structure feature comprises:

[0029] The global attention mechanism of the target detection model is used to perform feature extraction on the enhanced image, and the target region is collectively predicted in combination with the target structure feature to output the element category, the bounding box coordinate, and the confidence of the switchgear equipment corresponding to the target region.

[0030] In a second aspect, the present application further provides a switchgear equipment image detection system, comprising:

[0031] A data acquisition module is configured to acquire a to-be-detected image of the switchgear equipment.

[0032] An enhancement processing module is configured to input the to-be-detected image into a pre-trained image enhancement model to perform enhancement processing on a target structure feature in the to-be-detected image, and obtain an enhanced image.

[0033] The image enhancement model is a model pre-trained based on imaging environment parameters of the switchgear room.

[0034] A target detection module is configured to input the enhanced image into a target detection model to identify and output a target region corresponding to the target structure feature.

[0035] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0036] Acquiring a to-be-detected image of the switchgear equipment.

[0037] input the to-be-detected image into a pre-trained image enhancement model, perform enhancement processing on a target structural feature in the to-be-detected image, and obtain an enhanced image;

[0038] The image enhancement model is a model pre-trained based on imaging environment parameters of the power distribution room.

[0039] The enhanced image is input into a target detection model, and a target region corresponding to the target structural feature is identified and output.

[0040] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the following steps:

[0041] obtaining a to-be-detected image of the power distribution room equipment;

[0042] inputting the to-be-detected image into a pre-trained image enhancement model, performing enhancement processing on a target structural feature in the to-be-detected image, and obtaining an enhanced image;

[0043] The image enhancement model is a model pre-trained based on imaging environment parameters of the power distribution room.

[0044] The enhanced image is input into a target detection model, and a target region corresponding to the target structural feature is identified and output.

[0045] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the following steps:

[0046] obtaining a to-be-detected image of the power distribution room equipment;

[0047] inputting the to-be-detected image into a pre-trained image enhancement model, performing enhancement processing on a target structural feature in the to-be-detected image, and obtaining an enhanced image;

[0048] The image enhancement model is a model pre-trained based on imaging environment parameters of the power distribution room.

[0049] The enhanced image is input into a target detection model, and a target region corresponding to the target structural feature is identified and output.

[0050] The power distribution room equipment image detection method, device, computer device, computer readable storage medium and computer program product solve the problems of illumination and environmental interference, pre-train an image enhancement model based on actual imaging environment parameters of the power distribution room, accurately adapt to complex lighting conditions such as low light, backlight and reflection of the power distribution room, and interference factors such as dust on the surface of equipment and environmental noise, effectively improve the problems of image detail blur and insufficient contrast in the traditional method through directional enhancement processing of target structural features, solve the problem of insufficient generalization ability of the traditional model, and significantly improve the recognition degree of the target region corresponding to the target structural feature in the image to be detected.

[0051] The pre-trained image enhancement model can focus on the target structural features of the power distribution room equipment for targeted enhancement, improve the detection accuracy for specific elements of the power distribution room equipment, avoid the blindness of the traditional simple enhancement algorithm, make the target region in the enhanced image more clear and accurate, provide high-quality feature input for the subsequent target detection model, and solve the problem of information loss or interference superposition in the feature transmission process in the traditional technology. In addition, through the cooperative matching of image enhancement and target detection, the defects of independent work in the traditional technology are effectively made up, the missed detection and false detection caused by feature blur and interference are reduced, the accuracy and stability of the target region identification of the power distribution room equipment are significantly improved, and the engineering requirements of 24-hour uninterrupted intelligent monitoring of the power distribution room are met. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other related drawings without creative labor based on these drawings.

[0053] Figure 1 An application environment diagram of the power distribution room equipment image detection method in an embodiment;

[0054] Figure 2 A flowchart of the power distribution room equipment image detection method in an embodiment;

[0055] Figure 3 A before-and-after comparison diagram of image interference processing in an embodiment;

[0056] Figure 4 A flowchart of image enhancement processing of the power distribution room equipment in an embodiment;

[0057] Figure 5 A before-and-after comparison diagram of image enhancement processing in an embodiment;

[0058] Figure 6 A flowchart of a power distribution room equipment target detection method in an embodiment is shown in FIG. 1.

[0059] Figure 7 A result diagram of a power distribution room equipment target detection method in an embodiment is shown in FIG. 2.

[0060] Figure 8 A structural diagram of a power distribution room equipment image detection system in an embodiment is shown in FIG. 3.

[0061] Figure 9 An internal structure diagram of a computer device in an embodiment is shown in FIG. 4. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0063] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through a central element. In addition, "connected" in the following embodiments should be understood as "electrically connected", "communicatively connected" and the like if there is transmission of electrical signals or data between the connected objects.

[0064] Due to low-light environment, the target structure features of the to-be-detected image, such as low contrast and blurred details of the circuit breaker contact and the instrument scale, are prone to be missed or misjudged in traditional detection due to weak feature signals. At the same time, strong noise environment, such as image noise points caused by dust covering on the equipment surface, light reflection spots, and electromagnetic interference, can mask the target structure features, and the noise is prone to be misjudged as features in traditional detection, or the features are missed due to being blocked, and the noise interference can destroy the integrity of the feature boundary, affecting the positioning accuracy.

[0065] The present application proposes a power distribution room equipment image detection method, which performs scene adaptation pre-training and channel attention mechanism reinforcement on an image enhancement model, solves the feature attenuation and noise interference problems in low-light / strong-noise environment from the source, provides high-recognizability target structure features for the target detection model, and maximizes the advantages of feature matching and region positioning based on the enhanced high-quality features, so that the target structure features in low-light / strong-noise environment can be accurately recognized and locked, and finally realizes the stability and reliability of the power distribution room equipment detection in complex environment.

[0066] The power distribution room equipment image detection method provided by the embodiments of the present application can be applied to, for example, Figure 1The application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The terminal 102 can be a terminal device with data acquisition and local lightweight processing, including but not limited to industrial cameras, intelligent inspection robots, mobile inspection terminals (such as tablets / cell phones equipped with high-definition cameras), power distribution room fixed monitoring terminals, etc., used to collect power distribution room equipment images to be detected, receive server returned target area recognition results or trigger power distribution room intelligent control system to trigger early warning. The server 104 can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services, used for model deployment and core data processing, and needs to have a certain GPU computing power to support model inference. Those skilled in the art will think of many other variations, which are not limited by the present application.

[0067] Specifically, after the terminal 102 collects the image to be detected, it transmits the image data to the server 104 through wired or wireless networks; the server 104 uses the pre-trained model to complete the core processing of image enhancement and target detection, avoiding processing delays caused by insufficient terminal computing power; the server 104 returns the lightweight recognition result to the terminal 102, feeding back the recognition result such as target area coordinates, defect labeling, etc. In addition, if an edge computing server is deployed, the data transmission distance can be further shortened, the detection response speed in complex scenarios can be improved, and the real-time monitoring needs of the power distribution room can be adapted.

[0068] The power distribution room equipment image detection method of the present application can include a training phase and a test phase. In the training phase, the power distribution room imaging environment parameters and a large number of power distribution room equipment original images are collected, a training data set containing different interference scenes is constructed, and the target structure features in the data set are labeled, an image enhancement model is constructed and the image enhancement model is trained using the training data set to obtain a trained image enhancement model. The training process can be deployed on the server 104, and the terminal 102 sends the collected training data set content to the server 104, which trains the image enhancement model, so that the model can focus on the information of the target structure feature related channel.

[0069] After the trained image enhancement model is obtained, during the testing phase, terminal 102 can use its image acquisition device to capture images of the power distribution room equipment, upload them to server 104, and execute the power distribution room equipment image detection method of this application embodiment. First, the image is enhanced, and then the enhancement result is input into the target detection model to obtain the recognition result, which is then sent back to terminal 102. It is understood that the execution location of some and all steps of the above power distribution room equipment image detection method can be appropriately adjusted between terminal 102 and server 104. For example, the image enhancement model can be trained on server 104, and after training is completed and the trained image enhancement model is obtained, it can be deployed from server 104 to terminal 102 for use, etc. This application does not limit this.

[0070] In one exemplary embodiment, such as Figure 2 As shown, a method for image detection of equipment in a power distribution room is provided. This embodiment uses the application of this method to a server as an example for illustration. During the testing phase, it includes the following steps 210 to 230. Wherein:

[0071] Step 210: Obtain the image of the equipment to be inspected in the power distribution room.

[0072] Among them, power distribution room equipment refers to key equipment in the power distribution link of the power system, specifically including transformers, circuit breakers, disconnect switches, instruments, switch cabinets, and other equipment used for power transmission, distribution, and control. The image to be detected refers to the on-site image of the industrial equipment, specifically the original image captured on-site in the power distribution room by the image acquisition device in terminal 102, containing the components of the power distribution room equipment to be identified. Due to the enclosed environment of the power distribution room, there may be problems such as low light, backlight, reflection, and dust interference. The original image usually shows its grayscale histogram concentrated in the low brightness range, and due to the gain amplification of the sensor under low light, it is accompanied by Gaussian noise or salt-and-pepper noise.

[0073] For example, an image acquisition terminal is deployed to a designated location in the power distribution room. The terminal 102 needs to be adapted to the complex environment of the power distribution room with low light and high interference. According to the preset inspection path or manual triggering command, it takes pictures of the equipment in the power distribution room and acquires the original image containing the target equipment as the image to be detected. The terminal performs lightweight preprocessing on the acquired image to be detected, such as converting the format to JPEG (Joint Photographic Experts Group, JPG) / PNG (Portable Network Graphics, PNG), standardizing the resolution, and performing basic compression to avoid data redundancy. The preprocessed image to be detected is transmitted to the server 104 through a wired (Ethernet) or wireless (5G, Wi-Fi) network to prepare for subsequent enhancement processing.

[0074] Specifically, in one embodiment, a wheeled inspection robot is used to move to a designated position in front of the power distribution cabinet according to the magnetic track pasted on the ground, to ensure that the vertical distance from the camera lens to the power distribution cabinet is within the error allowed range, and then the monocular visible light area array camera carried by the robot starts image collection work, and presents real-time high-quality images to the connected terminal device through software.

[0075] Step 220, input the image to be detected into the pre-trained image enhancement model, enhance the target structure features in the image to be detected, and obtain an enhanced image.

[0076] Wherein, the target structure features refer to the key identification features of the power distribution room equipment, and specifically refer to the core structure information of the power distribution room equipment for distinguishing equipment types and judging running states, such as the position of the transfer switch, the switch contact form of the circuit breaker, the scale identification of the instrument, the arrangement mode of the wiring terminal, etc. After the image to be detected is processed by the pre-trained image enhancement model, an enhanced image with improved target structure feature recognition and suppressed environmental interference is obtained.

[0077] Illustratively, the server 104 receives the image to be detected transmitted by the terminal 102, inputs it into the pre-trained image enhancement model; the model automatically identifies the target structure feature related information in the image to be detected through the trained feature extraction logic, strengthens the signal strength of this type of features using the channel attention mechanism, and suppresses the noise signals caused by low light and electromagnetic interference; after the model processing, an enhanced image with clear target structure features, moderate contrast, and weakened interference factors is output.

[0078] Step 230, input the enhanced image into the target detection model, identify and output the target region corresponding to the target structure features.

[0079] Wherein, the target detection model can be an industrial target recognition model, which is an intelligent algorithm model with equipment structure feature recognition capability, such as improved models like YOLO, Faster R-CNN, DETR, etc. It can locate and identify the region corresponding to the target structure features from the enhanced image. The target region is the equipment feature positioning region, which specifically refers to the image region output by the target detection model and accurately matched with the target structure features, usually presented in the form of coordinate frame, mask, feature type, etc., containing key information such as region position, range, etc.

[0080] Exemplarily, the server 104 inputs the enhanced image obtained in step 220 into a deployed target detection model that has been trained in advance with samples of target structural features of equipment in the power distribution room and has the ability to recognize such features. The target detection model extracts features from the enhanced image and matches the extracted features with a pre-set target structural feature template in the model. After a successful match, the model determines the specific position and range of the target structural feature in the image through coordinate calculation and generates a corresponding target region, for example, labeled with a rectangular coordinate frame. The server 104 formats the target region information and transmits it back to the terminal 102 device through the network or stores it in the database for staff to view and subsequent analysis.

[0081] In the above power distribution room equipment image detection method, the enhanced image output by the image enhancement model is precisely adapted to the input requirements of the target detection model, forming a closed-loop collaborative link from feature optimization to feature recognition, and specifically addressing the detection pain points in low-light / high-noise environments.

[0082] Specifically, in a low-light environment, the image enhancement model is pre-trained with low-light parameters for the power distribution room. It can automatically identify the weak signals of target structural features in low-light images, focus on the channel information corresponding to such features through a channel attention mechanism, and strengthen the gray difference between the features and the background, for example, to improve the contrast of instrument scales and dials and highlight the edge contours of contacts. At the same time, it suppresses the overall noise caused by low light. The final output enhanced image converts the originally blurred target structural features into clear and distinguishable high-recognition features. Further, when the target detection model receives the enhanced high-recognition features, it does not need to deal with the problem of feature attenuation caused by low light and can directly perform precise matching with the pre-trained feature template. Due to the improvement in feature signal strength, the mis-matching rate in the matching process is significantly reduced, thereby achieving high-accuracy recognition of target structural features. At the same time, the clear feature edges provide a clear boundary reference for the model, allowing it to accurately lock the feature range during coordinate calculation and avoid positioning deviation caused by fuzzy edges.

[0083] In a strong noise environment, the image enhancement model is pre-trained based on common noise types in the power distribution room. The noise recognition logic formed by interference simulation training can accurately distinguish target structure features from environmental noise. Through channel attention mechanism, it can suppress invalid channel signals corresponding to noise, such as highlight channels of reflected light spots and low gray channels covered by dust. At the same time, it can retain and strengthen the effective channel information of the target structure features, such as the arrangement channel of the terminal and the outline channel of the switch of the power distribution cabinet, so as to realize the synchronization of noise reduction and feature enhancement, and ensure the purity and integrity of the target features in the enhanced image. Then, when the enhanced image input by the target detection model has eliminated most of the noise interference, only the effective target features need to be focused on in the feature template matching process, avoiding the confusion between noise and features, and significantly reducing the false detection rate. At the same time, the complete feature boundary provides accurate positioning basis for the model, and the coordinate frame can strictly fit the actual range of the feature, solving the positioning deviation problem caused by noise shielding and realizing high positioning accuracy.

[0084] In an exemplary embodiment, in the training stage, the pre-trained image enhancement model is trained based on specific imaging environment parameters of the power distribution room, such as illumination intensity, interference type and device layout, through a labeled training data set. The model has the ability to directionally enhance the target structure features of the power distribution room equipment, and the core adopts a low-light image enhancement network that integrates channel attention mechanism.

[0085] For example, a large number of original images of power distribution room equipment are obtained, and environmental parameters that explicitly affect image quality, such as low-light threshold, common interference types and device layout density, are used to simulate low-light, reflection, dust coverage and electromagnetic noise through algorithms to construct a training data set containing different interference scenarios, and label the target structure features in the data set. An image enhancement model with a low-light image enhancement network as the basic framework is constructed, and a channel attention mechanism is integrated into the feature extraction network of the model to enable the model to focus on the information of the target structure feature-related channels. The model is trained using the training data set, and the model parameters are adjusted through back propagation to enhance the model's response to target structure features and suppress the influence of interference factors. Finally, a pre-trained image enhancement model is obtained and deployed to the server 104.

[0086] Specifically, in one embodiment, first, all input original images are normalized according to a unified size and stored in a standardized format (such as JPEG or PNG) for subsequent processing. Second, using Python language, combined with random geometric transformation and smoke noise superposition technology, the original images are processed into a large-scale image data set that can simulate multiple situations, such as Figure 3 As shown in the left image of the figure, it is an original image, and the right image is an image after interference processing.

[0087] The training data set includes training images obtained by performing interference processing based on the imaging environment parameters of the power distribution room, and the training images are generated in the following manner:

[0088] According to the imaging environment parameters of the power distribution room, a brightness attenuation threshold and an environmental medium concentration range are set; wherein the brightness attenuation threshold is determined by measuring the illumination intensity of the power distribution room at different time periods using a luxmeter, and the statistical distribution of insufficient illumination is determined, and the environmental medium concentration range is simulated to simulate the scattering effect of air dust or interference noise on light.

[0089] An original image of the power distribution room equipment is obtained, random geometric transformation is performed on the original image, and a low-illumination feature and a smoke noise layer are superimposed on the original image according to the brightness attenuation threshold and the environmental medium concentration range, to generate a training image.

[0090] Wherein, the random geometric transformation is used to simulate the image form difference caused by the change of camera angle and distance in actual inspection, so as to improve the recognition ability of the model to different poses of the target structure feature. According to the above brightness attenuation threshold, the overall gray value of the image is adjusted by algorithm: the RGB channel value of each pixel of the image is multiplied by the attenuation coefficient, and the channel contrast is reduced at the same time, and the weak gray difference is retained for the target structure feature area, so that the model can learn the core outline of the feature under low light during training. Verify whether the image brightness after superposition reaches the threshold range, and adjust the attenuation coefficient again if it does not meet the standard. Based on the environmental medium concentration range, for example, the concentration level is divided into low, medium and high, and the gray offset is 5-15, 16-30 and 31-50 respectively. A semi-transparent smoke layer is generated by a random noise generation algorithm, the size of the layer is consistent with the original image, and an image weighted fusion algorithm is used to superimpose the smoke noise layer and the image processed by low-illumination, so that the noise naturally covers the image background and the target area, and the real interference shielding effect is simulated.

[0091] Specifically, the optional steps of image enhancement are as follows:

[0092] (1) Random rotation: the image enhancement ImageEnhance module in the image processing library (such as Python Imaging Library, abbreviated as PIL library) is used to realize random angle rotation, and the angle range is set to ±15 degrees to ensure that the enhanced image remains identifiable. By simulating image features at different angles, the rotation invariance of the model is improved.

[0093] (2) Random flip: the image transformation ImageTransform module in the image processing library is used to realize image transformation, and random selection is made among horizontal flip, vertical flip and no flip, which can increase the diversity of the data set and cover more possible scene views.

[0094] (3) Random Cropping: Implement image cropping using the ImageOps module in the image processing library. Randomly select a cropping ratio (80%~100%) while maintaining the original aspect ratio. This method simulates different regions of interest and avoids overfitting global features.

[0095] (4) Random Brightness Adjustment: Use the ImageEnhance module in the image processing library to adjust brightness. Set the brightness range to 50%~100% of the original brightness. This simulates image features under different lighting conditions and avoids the problem of over-enhancement leading to unrecognizable images.

[0096] (5) Random Contrast Adjustment: Use the ImageEnhance module in the image processing library to adjust contrast. Set the contrast range to 50%~150% of the original contrast. This avoids the problem of low contrast leading to feature loss or high contrast leading to overfitting. It increases the amount of visual information in the image and improves the model's ability to learn detailed features.

[0097] (6) Smoke Effect Superposition: Use the computer vision library (such as OpenCV library) to generate a random noise image and combine it with the original image using the image processing library. Use the cv2.randn() function in the computer vision library to generate Gaussian noise. The noise intensity is controlled by adjusting the standard deviation (such as 10). The transparency of the noise layer is set as a adjustable parameter, for example, randomly selected between 0.1 and 0.5, to simulate different concentrations of smoke effects.

[0098] To achieve non-uniform smoke effects (such as some areas being denser and some areas being lighter), first generate a random grayscale image on a blank image, which represents the distribution of smoke concentration. Use cv2.GaussianBlur() to perform Gaussian blur on the grayscale image to simulate the diffusion characteristics of smoke. Use the blurred grayscale image as the basis for the non-uniform smoke layer and combine it with the original noise layer.

[0099] To enhance the realism and level of smoke, multiple smoke layers with different concentrations and sizes can be superimposed on the image. For example, by randomly adjusting the transparency, size, and position of each smoke layer. After each superposition, use the cv2.addWeighted() function to achieve the fusion effect.

[0100] The key parameters for generating smoke include: the standard deviation of the noise (controls the intensity of the noise), the kernel size of the Gaussian blur (controls the diffusion degree of the smoke), the number of multi-layer superposition (such as 1~3 times), and the transparency of each layer. By adjusting the above parameters, you can get reasonable and realistic images.

[0101] The image processed by the above method is saved to a training data set folder, and the training image is manually or automatically sampled to ensure that the low-light feature, the smoke noise layer is superimposed naturally, the target structure feature is not completely covered, and the invalid image is removed to ensure the quality of the training data set. Through the above process, the complex environment of low light and high noise is simulated to enhance the adaptability of the model to the real scene.

[0102] In one exemplary embodiment, as shown in Figure 4 The step 220 of enhancing the target structure feature in the image to be detected includes steps 410 to 440. Among them:

[0103] Step 410, the image enhancement model adopts a low-light image enhancement network fused with a channel attention mechanism, and uses the low-light image enhancement network to decompose the image to be detected into a light component and a reflection component.

[0104] The light component represents the ambient light, which is a disturbance factor, and the reflection component represents the essential attributes of the power distribution cabinet components, such as color, texture, and edge, which are the basis for judgment. Although the traditional low-light image enhancement network can complete the image enhancement process, it has problems such as incomplete denoising and missing details, so the model is improved in this embodiment, and a channel attention mechanism is introduced into the backbone network to obtain better image enhancement effect. The channel attention mechanism can help the model to accurately identify the details in the image, and can make the model focus on the channel features corresponding to some regions, thereby enhancing the clarity and contrast of the details.

[0105] Step 420, the channel attention mechanism is embedded in the backbone network for processing the reflection component, and the feature map of the reflection component is extracted through the convolution layer of the backbone network. The model extracts the feature map of the reflection component through the convolution layer. At this time, the feature map contains multiple channels, some channels respond to the edges of high-frequency information components, and some channels respond to background noise or false textures under low light.

[0106] Step 430, the feature map contains multiple channels, the channel presenting local high frequency and high contrast is determined as the target channel where the target structure feature is located, the weight of the target channel is increased, and the weighted reflection component is formed.

[0107] Step 440, the weighted reflection component and the light component are combined to reconstruct an enhanced image. The weighted reflection component features are used to combine the light component after gamma correction to reconstruct an enhanced image. After the image enhancement model processing, the dark noise in the image to be detected is suppressed by the attention mechanism, and the device edge is retained. The enhanced image obtained has a significantly higher signal-to-noise ratio than the image to be detected while improving the overall brightness.

[0108] In this embodiment, as shown in Figure 5As shown, it is a pair of contrast images of image enhancement, the left image is before enhancement, and the right image is after enhancement. Through feature extraction and attention weighting, the pixel value distribution of the device can be stretched to the normal gray scale range, and the gradient value of the edge of the component is significantly increased, and the variance of the noise area is significantly reduced, while keeping the device geometric position unchanged and the spatial consistency of the image to be detected. Decoupling and balancing of denoising and enhancement are achieved, avoiding the over-enhancement phenomenon of full-screen noise caused by traditional methods to see the device, ensuring that the features extracted in the subsequent detection step are real device textures rather than false noise textures, thereby reducing the false positive rate.

[0109] Further, the channel attention mechanism is embedded in the backbone network processing the reflection component in step 420, and the feature map of the reflection component is extracted through the convolution layer of the backbone network, including:

[0110] The channel attention mechanism adopts an efficient channel attention network module. Traditional channel attention (such as SE-Net) needs to reduce dimension through a fully connected layer, which will lose part of the channel information and increase the parameter quantity. Through the one-dimensional convolution kernel of the efficient channel attention network module, the information interaction of adjacent channels in the feature map is calculated to obtain the associated channel feature vector; the one-dimensional convolution only considers the information interaction of the current channel and its adjacent k channels. The convolution kernel size k is adaptively selected according to the channel dimension. The associated channel feature vector is activated by a compression function to generate a channel weight vector; the channel weight vector and the feature map are multiplied by the channel dimension to obtain the weighted reflection component. The feature vector of the input dimension can be output with an importance label through extremely small calculation cost.

[0111] Through this method, the computational complexity and parameter quantity of the model can be significantly reduced, while the channel dimension information is not lost. This algorithm can be deployed on a low-power embedded inspection camera in a power distribution room or a mobile inspection robot with limited computing power terminal 102, to realize real-time frame-by-frame enhancement and detection of video stream, not just processing static pictures, solving the high requirement of real-time detection in industrial sites.

[0112] Further, the target structure feature in step 430 is a local high-frequency, high-contrast, and specific direction texture feature in the feature map of the reflection component.

[0113] Wherein, the feature map of the reflection component is processed by a high-frequency filtering algorithm to detect the region of rapid change of local gray value, and the candidate region of the target structure feature is preliminarily locked; the gray contrast of the candidate region is calculated, and the region with a contrast higher than a preset threshold (such as a difference ≥ 50, which is set based on training data statistics) is screened out to eliminate the fuzzy region with low contrast; the high-contrast candidate region is filtered by a direction filter to match the texture feature conforming to a specific direction, and the target structure feature is finally determined; the corresponding feature channel of the target structure feature in the reflection component feature map is located, and the channel index is recorded for subsequent weight adjustment.

[0114] The channel attention mechanism increases the weight of the target channel where the target structure feature is located, and reduces the weight of the channel corresponding to the false texture introduced by low illumination and noise:

[0115] The channel attention mechanism module adaptively evaluates all channels of the reflection component feature map, calculates the feature response intensity of each channel, assigns a higher weight to the target channel marked in the above step according to its feature response intensity, for example, multiplies the original weight by a reinforcement coefficient of 1.5-2.0 to amplify the signal intensity of the target structure feature, so that it is more prominent in the feature map; through a noise recognition algorithm, for example, by counting the texture regularity of each channel, the channel with irregular and chaotic fluctuations is determined as a false texture channel, and the channel corresponding to the false texture introduced by low illumination and noise is located; a low weight is assigned to the false texture channel, for example, the original weight is multiplied by an inhibition coefficient of 0.1-0.3 to weaken the pseudo-feature signal carried by it and reduce the interference with the target structure feature; the adjusted weights of all channels are normalized to map to the interval of 0-1, ensuring the balance of the weights between channels and avoiding feature distortion caused by excessive reinforcement or inhibition.

[0116] In an exemplary embodiment, as shown in Figure 6 Fig. 2, step 230 inputs the enhanced image into the target detection model to identify and output the target region corresponding to the target structure feature, including steps 610-630. Among them:

[0117] Step 610: The global attention mechanism of the target detection model is used to extract features from the enhanced image. The global attention mechanism means that the model not only pays attention to the pixels inside a certain candidate box, but also calculates the correlation between the box and other regions of the whole image. The convolution kernel of the detection model slides on the enhanced image. Due to the enhancement in the previous step, the rectangular or circular edge profile of the power distribution cabinet components will produce a strong activation response in the feature map. At the same time, the global attention mechanism analyzes the context, for example, when the cabinet body of the power distribution cabinet is detected as a feature, it will increase the probability prediction of the appearance of switches or instrument panels in its internal region, or indicate the global semantic regularity that the indicator lights usually appear in rows.

[0118] At step 620, the target region is collectively predicted in combination with the target structure features. According to the shape, texture, and color feature vectors extracted from the feature map, the feature vectors are mapped to class probabilities through a multi-classification activation function, and the coordinate offset of the target is calculated through a bounding box regression. When the local features are fuzzy (for example, two switches look very similar), the model uses the context information extracted by the global attention as a supplementary criterion.

[0119] At step 630, the element category of the power distribution cabinet equipment corresponding to the target region, the bounding box coordinates, and the confidence are output. The element category is a voltmeter, a closing button, etc., the bounding box coordinates accurately locate the position of the equipment in the image, and the confidence represents the determination degree of the model on the judgment.

[0120] In an exemplary embodiment, the target detection model is preferably an end-to-end target detection model (DEtection TRansformer, DETR) based on a Transformer architecture, which realizes target recognition through a global attention mechanism and collective prediction without the need for pre-defined anchor boxes. However, the model lacks attention to local key features and is easily disturbed by the background. The image enhancement module in the preceding step performs channel screening on the feature map in advance, strengthens the core features such as the edges and textures of the power distribution cabinet components, and then inputs the Transformer encoder of the DETR, which can make the global attention of the DETR not waste computing resources on invalid features, accurately match the strengthened component features, faster lock the target region, and avoid missing detection (such as small components: indicator light, terminal) or false detection (such as a reflective area being recognized as a component) due to the loss of local features.

[0121] It should be understood that the training data set of the target detection model can be manually labeled, and the element category and bounding box coordinates of the target structure features are labeled for the training image. The labeled data is input into the DETR model, and the model is trained to recognize and locate the target features.

[0122] By performing the operation of step 630, the target region where the transfer switch is located and its confidence can be obtained as shown in FIG. 6C. Figure 7

[0123] The embodiment fully utilizes the strengthened target structure features in the enhanced image through the global attention mechanism of the end-to-end target detection model, realizes integrated processing from feature extraction to target matching to result output, forms a collaborative closed loop from feature enhancement to accurate recognition with the image enhancement model, and specifically improves the detection accuracy and positioning precision in a low-light / strong-noise environment.

[0124] ​It should be understood that although each step in the flowchart involved in the above embodiments is shown in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above embodiments can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps. It can be understood that the steps in different embodiments can be freely combined as needed, and various non-contradictory schemes formed by the combination are within the scope of protection of the present application.

[0125] Based on the same inventive concept, the embodiments of the present application also provide a power distribution room equipment image detection system for implementing the above-mentioned power distribution room equipment image detection method. The problem-solving implementation scheme provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more power distribution room equipment image detection system embodiments provided below can refer to the limitations of the power distribution room equipment image detection method in the above text, which will not be repeated here.

[0126] In an exemplary embodiment, as shown in Figure 8 a power distribution room equipment image detection system is provided, comprising: a data acquisition module 810, an enhancement processing module 820, and a target detection module 830, wherein:

[0127] The data acquisition module 810 is configured to acquire an image to be detected of the power distribution room equipment. It can be a fixed acquisition device, such as a high-definition network camera or a binocular camera deployed at a key point of the power distribution room, connected to a local area network switch of the power distribution room through a network cable or an optical fiber. It can also be a mobile acquisition device, such as a track inspection robot, a wheeled inspection robot, or a drone, etc. In this embodiment, a wheeled inspection robot with a high-definition camera mounted thereon is preferred, which communicates with a base station through a Wi-Fi, 4G / 5G private network, or ZigBee wireless communication module, and then accesses the local area network.

[0128] The enhancement processing module 820 is configured to input the image to be detected into a pre-trained image enhancement model, perform enhancement processing on the target structure features in the image to be detected, and obtain an enhanced image; wherein the image enhancement model is a model pre-trained based on the imaging environment parameters of the power distribution room.

[0129] The target detection module 830 is configured to input the enhanced image into a target detection model, identify and output a target region corresponding to the target structure features.

[0130] As the core module of the application, the running physical carrier of the enhancement processing module 820 and the target detection module 830 can be an industrial control computer or a high-performance image processing unit server deployed in the local cabinet of the power distribution room, or an embedded edge computing gateway. The processor thereof receives the original video stream or picture transmitted by the collection terminal, performs image enhancement processing and target detection in real time, and outputs the device state result. Through the internal bus or network interface, the processed structured data, such as No. 1 cabinet-circuit breaker-opening-confidence 0.99, is sent to the monitoring client or display terminal.

[0131] The monitoring client / display terminal can be a PC workstation or a spliced large screen in the power distribution room duty room, or a tablet computer or smart phone in the hands of the operation and maintenance personnel. The enhanced clear image and the target detection bounding box are visually superimposed and displayed, and an alarm is issued for abnormal states (such as abnormal meter display, abnormal transfer switch position, and circuit breaker unable to close or open).

[0132] The image enhancement model in the enhancement processing module 820 is trained using a training data set, which includes training images obtained by performing interference processing based on the imaging environment parameters of the power distribution room. The generation method of the training images includes:

[0133] According to the imaging environment parameters of the power distribution room, set the brightness attenuation threshold and the environmental medium concentration range;

[0134] Obtain the original image of the power distribution room device, perform random geometric transformation on the original image, and superimpose a low-illumination feature and a smoke noise layer on the original image according to the brightness attenuation threshold and the environmental medium concentration range, to generate a training image.

[0135] Further, the image enhancement model adopts a low-light image enhancement network that fuses a channel attention mechanism;

[0136] The target structure feature in the to-be-detected image is enhanced, including:

[0137] The to-be-detected image is decomposed into an illumination component and a reflection component using the low-light image enhancement network;

[0138] The channel attention mechanism is embedded in the backbone network that processes the reflection component, and the feature map of the reflection component is extracted through the convolution layer of the backbone network;

[0139] The feature map contains multiple channels, and the channel that presents local high frequency and high contrast is determined as the target channel where the target structure feature is located, the weight of the target channel is increased, and a weighted reflection component is formed;

[0140] The weighted reflection component is combined with the illumination component to reconstruct an enhanced image.

[0141] Further, the channel attention mechanism adopts an efficient channel attention network module:

[0142] The channel attention mechanism is embedded in the backbone network for processing the reflection component, and the feature map of the reflection component is extracted through the convolution layer of the backbone network, including:

[0143] The information interaction of adjacent channels in the feature map is calculated through the one-dimensional convolution kernel of the efficient channel attention network module, and a correlation channel feature vector is obtained;

[0144] The correlation channel feature vector is activated by a compression function to generate a channel weight vector;

[0145] The channel weight vector is multiplied with the feature map in the channel dimension to obtain a weighted reflection component.

[0146] Further, the target structure feature is a texture feature that is locally high-frequency, high-contrast and specific direction in the feature map of the reflection component;

[0147] The channel attention mechanism increases the weight of the target channel where the target structure feature is located, and reduces the weight of the channel corresponding to the false texture introduced by low illumination and noise.

[0148] In the target detection module 830, the target detection model is an end-to-end target detection model trained by labeled data;

[0149] The enhanced image is input into the target detection model, and the target region corresponding to the target structure feature is identified and output, including:

[0150] The enhanced image is feature-extracted by the global attention mechanism of the target detection model, and the target region is collectively predicted in combination with the target structure feature, and the element category, the bounding box coordinates and the confidence of the power distribution cabinet equipment corresponding to the target region are output.

[0151] Each module in the above power distribution room equipment image detection system can be realized by software, hardware and their combinations. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations of the above modules.

[0152] In an exemplary embodiment, a computer device is provided, which can be a terminal, and its internal structure diagram can be as shown in Figure 9As shown in the figure. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be realized through WIFI, mobile cellular network, Near Field Communication (NFC) or other technologies. The computer program is executed by the processor to realize a power distribution room equipment image detection method. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device.

[0153] Those skilled in the art can understand that, Figure 9 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0154] In one exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps:

[0155] Obtaining a to-be-detected image of the power distribution room equipment;

[0156] Inputting the to-be-detected image into a pre-trained image enhancement model to perform enhancement processing on the target structure features in the to-be-detected image and obtaining an enhanced image; wherein the image enhancement model is a model pre-trained based on the imaging environment parameters of the power distribution room;

[0157] Inputting the enhanced image into a target detection model to identify and output the target region corresponding to the target structure features.

[0158] In one embodiment, the processor executing the computer program further implements the following steps:

[0159] The training data set includes training images obtained by performing interference processing based on the imaging environment parameters of the power distribution room, and the generation manner of the training images includes:

[0160] According to the imaging environment parameters of the power distribution room, a brightness attenuation threshold and an environmental medium concentration range are set;

[0161] An original image of the power distribution room equipment is acquired, random geometric transformation is performed on the original image, and a low-illumination feature and a smoke noise layer are superimposed on the original image according to the brightness attenuation threshold and the environmental medium concentration range, to generate a training image.

[0162] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0163] The image enhancement model adopts a low-illumination image enhancement network that fuses a channel attention mechanism;

[0164] The target structural features in the to-be-detected image are enhanced, including:

[0165] The to-be-detected image is decomposed into an illumination component and a reflection component by using the low-illumination image enhancement network;

[0166] The channel attention mechanism is embedded in the backbone network for processing the reflection component, and a feature map of the reflection component is extracted by a convolution layer of the backbone network;

[0167] The feature map contains multiple channels, a channel that presents local high frequency and high contrast is determined as a target channel where the target structural features are located, the weight of the target channel is increased, and a weighted reflection component is formed;

[0168] The weighted reflection component is combined with the illumination component to reconstruct and generate an enhanced image.

[0169] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0170] The channel attention mechanism adopts an efficient channel attention network module:

[0171] The channel attention mechanism is embedded in the backbone network for processing the reflection component, and a feature map of the reflection component is extracted by a convolution layer of the backbone network, including:

[0172] An information interaction between adjacent channels in the feature map is calculated by using a one-dimensional convolution kernel of the efficient channel attention network module, to obtain a correlation channel feature vector;

[0173] The correlation channel feature vector is activated by a compression function to generate a channel weight vector;

[0174] The channel weight vector is multiplied with the feature map in the channel dimension to obtain a weighted reflection component.

[0175] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0176] The target structure feature is a local high-frequency, high-contrast, and specific-direction texture feature in a feature map of a reflection component;

[0177] The channel attention mechanism increases the weight of a target channel where the target structure feature is located, and reduces the weight of a channel corresponding to a false texture introduced by low illumination and noise.

[0178] In an embodiment, the processor, when executing the computer program, also implements the following steps:

[0179] The target detection model is an end-to-end target detection model trained by labeled data.

[0180] The enhanced image is input into the target detection model to identify and output a target region corresponding to the target structure feature, including:

[0181] The global attention mechanism of the target detection model is used to extract features of the enhanced image, and the target region is collectively predicted in combination with the target structure feature to output an element category, a bounding box coordinate, and a confidence of the power distribution cabinet equipment corresponding to the target region.

[0182] In an embodiment, a computer-readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0183] In an embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0184] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use, and processing of related data need to comply with relevant regulations.

[0185] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0186] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0187] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for image detection of equipment in a power distribution room, characterized in that, The method includes: Acquire the image of the equipment in the power distribution room to be inspected; The image to be detected is input into a pre-trained image enhancement model to enhance the target structural features in the image to be detected, thereby obtaining an enhanced image. The image enhancement model mentioned above is a model pre-trained based on the imaging environment parameters of the power distribution room; The enhanced image is input into the target detection model, which identifies and outputs the target region corresponding to the target structural features.

2. The method according to claim 1, characterized in that, The image enhancement model is trained using a training dataset, which includes training images obtained by performing interference processing based on the imaging environment parameters of the power distribution room. The training images are generated in the following ways: Based on the imaging environment parameters of the power distribution room, set the brightness attenuation threshold and the environmental medium concentration range; The original image of the power distribution room equipment is acquired, a random geometric transformation is performed on the original image, and a low-light feature and a smoke noise layer are superimposed on the original image according to the brightness attenuation threshold and the ambient medium concentration range to generate the training image.

3. The method according to claim 1, characterized in that, The image enhancement model employs a low-light image enhancement network that incorporates a channel attention mechanism; The enhancement processing of the target structural features in the image to be detected includes: The low-light image enhancement network is used to decompose the image to be detected into an illumination component and a reflection component; The channel attention mechanism is embedded in the backbone network that processes the reflection component, and the feature map of the reflection component is extracted through the convolutional layer of the backbone network. The feature map contains multiple channels. The channel that presents local high frequency and high contrast is identified as the target channel where the target structural feature is located. The weight of the target channel is increased to form a weighted reflection component. The weighted reflection component is combined with the illumination component to reconstruct the enhanced image.

4. The method according to claim 3, characterized in that, The channel attention mechanism employs an efficient channel attention network module: The step of embedding the channel attention mechanism into the backbone network that processes the reflection component, and extracting the feature map of the reflection component through the convolutional layers of the backbone network, includes: The one-dimensional convolutional kernel of the efficient channel attention network module is used to calculate the information interaction between adjacent channels in the feature map to obtain the associated channel feature vector. The associated channel feature vectors are activated by a compression function to generate channel weight vectors; The channel weight vector is multiplied by the feature map along the channel dimension to obtain the weighted reflection component.

5. The method according to claim 3, characterized in that, The target structural features are local high-frequency, high-contrast, and specific-directional texture features in the feature map of the reflection component; The channel attention mechanism increases the weight of the target channel containing the target structural features and decreases the weight of the channels corresponding to false textures introduced by low illumination and noise.

6. The method according to claim 1, characterized in that, The target detection model is an end-to-end target detection model trained with labeled data; The step of inputting the enhanced image into the target detection model, identifying and outputting the target region corresponding to the target structural features includes: The enhanced image is feature extracted using the global attention mechanism of the target detection model. Combined with the target structural features, the target region is ensemble predicted, and the component category, bounding box coordinates, and confidence score of the power distribution cabinet equipment corresponding to the target region are output.

7. A power distribution room equipment image detection system, characterized in that, The system includes: The data acquisition module is used to acquire the image to be detected of the power distribution room equipment; The enhancement processing module is used to input the image to be detected into a pre-trained image enhancement model, and to enhance the target structural features in the image to be detected to obtain an enhanced image; The image enhancement model mentioned above is a model pre-trained based on the imaging environment parameters of the power distribution room; The target detection module is used to input the enhanced image into the target detection model, identify and output the target region corresponding to the target structural features.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images