An AI recognition-based power platform image data enhancement method

Abstract

The application discloses an electric power platform image data enhancement method based on AI recognition, relates to the field of image data enhancement, and comprises the following steps: constructing an electric power equipment image benchmark data set and training a multi-dimensional feature recognition model, extracting the structure, state and environment features of equipment to generate a structured descriptor; adaptively clustering images based on the descriptor to form subsets with similar feature attributes; designing matched and diversified image enhancement operation chains for the common features and defect types of each subset, generating synthesized images with consistent semantics but visual enhancement through parameter adaptive adjustment; and finally, pairing the synthesized images with the original images, filtering after verification to form an enhanced training data set. The application has the advantages that: relying on AI multi-dimensional feature recognition and adaptive image clustering, differentiated enhancement strategies are customized on demand and dynamically adjusted, complex electric power scenes and image defects are accurately adapted, and high-quality and diversified electric power image data is efficiently generated.

Description

Technical Field

[0001] This invention relates to the field of image data enhancement, and in particular to a method for enhancing image data of a power platform based on AI recognition. Background Technology

[0002] With the deepening of smart grid construction, computer vision-based power equipment inspection is becoming increasingly widespread. However, the training of its AI models heavily relies on massive amounts of high-quality labeled image data. In practical applications, the power field environment is complex, with a wide variety of equipment and scarce samples for some operating conditions. Furthermore, factors such as weather, obstructions, and shooting angles further complicate the process, making it costly and difficult to directly obtain sufficient, balanced, and high-quality field image data. This data bottleneck severely restricts the improvement of AI model recognition accuracy and generalization ability. Therefore, researching efficient data augmentation methods is crucial.

[0003] Current mainstream power image data augmentation methods suffer from significant technical shortcomings, lacking overall intelligence, adaptability, and professionalism. Existing methods largely employ traditional image processing techniques with fixed parameters, failing to consider the specific characteristics of power equipment and scene features. They lack multi-dimensional feature extraction and intelligent analysis capabilities, making it impossible to accurately identify differences in equipment structure, defect status, and environmental interference. These methods often perform global uniform augmentation without adaptive clustering or customized augmentation logic, failing to differentiate between various issues such as blurring, abnormal exposure, rain, snow, fog, haze, and equipment defects. This leads to problems like over-enhancement, loss of detail, weakening of defect features, and damage to the original semantic information of the image. Furthermore, traditional methods cannot achieve adaptive parameter adjustment, relying heavily on manual fixed configurations, resulting in extremely poor adaptability to complex power inspection scenarios. Most methods also lack robust realism verification and quality filtering mechanisms, leading to artifacts, distortion, and semantic shifts in the generated synthetic images. The datasets are of inconsistent quality and lack diversity, ultimately resulting in low training data effectiveness and significantly weakening the training effect and generalization ability of intelligent power equipment detection models in real-world scenarios. Summary of the Invention

[0004] To improve existing methods, this paper proposes an AI-based image data enhancement method for power platforms. This method relies on AI multi-dimensional feature recognition and adaptive image clustering to customize differentiated enhancement strategies and dynamically adjust parameters as needed. It can accurately adapt to various complex power scenarios and image defects, and efficiently generate high-quality and diverse power image data.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for enhancing image data of a power platform based on AI recognition, comprising: A raw image dataset from a power platform scenario is collected, and the raw image dataset is preprocessed and quality-screened to construct a benchmark image set. A multi-dimensional feature recognition model for power equipment is constructed and trained. The trained feature recognition model is used to analyze the benchmark image set, extract the structural features, state features and contextual features of the power equipment in each image, and generate corresponding structured feature descriptors. Based on the structured feature descriptor, the baseline image set is adaptively clustered to form multiple image subsets with similar feature attributes; For each subset of images, a diverse set of image enhancement operation chains are designed to match the common features and defect types represented by those features and defects. The image enhancement operation chains are composed of a series of ordered and parameter-adjustable low-level image processing operations. Within each of the image subsets, the parameters of the image enhancement operation chain are adaptively adjusted and the operation logic is selected based on the structured feature descriptor to generate a synthetic image that has the same semantic content as the original image but with enhanced visual representation. The synthesized images are paired with the corresponding original images, and after authenticity verification and quality filtering, they together form an enhanced image dataset for model training.

[0006] Preferably, the preprocessing and quality screening of the original image dataset to construct a benchmark image set specifically includes: By using fixed monitoring equipment, mobile inspection equipment or drones deployed at the power platform site, original images and video streams from multiple perspectives, time periods and weather conditions are collected, including power transmission and transformation equipment, lines, towers and insulators, and images are extracted from the video stream at preset time intervals. The acquired raw images are standardized, including: uniformly converting them to a preset color space and pixel resolution, performing optical distortion correction, and aligning the image horizon. An automatic filtering mechanism based on a sharpness evaluation function and information entropy assessment is adopted to remove invalid images that are severely blurry, overexposed, underexposed, or where key power equipment is largely obscured. Images with acceptable image quality are retained to form the reference image set, and each image in the reference image set is labeled with metadata tags indicating the acquisition time, equipment type, and weather conditions.

[0007] Preferably, the automatic filtering mechanism based on the sharpness evaluation function and information entropy assessment specifically includes: The input image is converted into a grayscale image, and the grayscale gradient magnitude and direction of each pixel in the image within its neighborhood are calculated to form a gradient map. Based on the gradient map, a comprehensive sharpness score is obtained by calculating the statistical characteristics of the gradient magnitude. By setting an empirical threshold, images with scores below this threshold are marked as severely blurry, classified as invalid images, and discarded. The image is divided into multiple regular grids, and the information entropy of the gray value distribution is calculated for each grid region. Based on the prior location of the power equipment, we focus on the grid information entropy of the area where the equipment is located. If the information entropy of the area is lower than the set threshold, the image details are severely insufficient, and the image is determined to be invalid and discarded. Analyze the global grayscale histogram of the image. If a large number of pixels in the histogram are concentrated at the high brightness end or the low brightness end, and the regional information entropy is extremely low, then it is determined to be an overexposed or underexposed image. Based on the completion of clarity and information entropy screening, the remaining images are segmented into the foreground region of power equipment. If the detected main equipment area is covered, it is determined that the key equipment is largely occluded and the image is removed. The filtered images are then compiled to form a baseline image set.

[0008] Preferably, the construction and training of the multi-dimensional feature recognition model for power equipment specifically includes: A deep convolutional neural network for multi-task learning is constructed as the backbone network of the feature recognition model. This backbone network shares the underlying feature extractor and derives three feature parsing branches: structural feature branch, state feature branch, and environment feature branch. The structural feature branches are used to identify the components, outlines, connection points and their spatial topological relationships of specific power equipment in the positioning image, and the structural integrity vector of the equipment is output. The physical state of the equipment surface is analyzed through the state feature branch, including but not limited to the detection and segmentation of rusted areas, oil stains, abnormal heat areas, cracks and damage, and the output of the equipment's state index vector. The overall contextual information of the image is analyzed through the environmental feature branch, including lighting conditions, fog concentration, rain and snow interference, background complexity, and distribution of occlusions, and an environmental factor vector is output. The feature recognition model is trained end-to-end using labeled power equipment image samples until the model stably and accurately outputs the structured feature descriptor, which is a high-dimensional tensor that fuses structural feature vectors, state feature vectors, and environmental feature vectors.

[0009] Preferably, the step of adaptively clustering the baseline image set based on the structured feature descriptor to form multiple image subsets with similar feature attributes specifically includes: Calculate the multidimensional feature distance between the structured feature descriptors of all images in the reference image set; A density-based clustering algorithm is used to automatically aggregate images with similar visual and semantic features into the same cluster based on the feature distance, and each cluster forms a subset of the images. For each of the image subsets, a core feature overview is summarized, which includes the dominant device type, the most common one or more defect states, and the most representative environmental interference type within the subset.

[0010] Preferably, the step of designing a diverse image enhancement operation chain that matches the common features and defect types represented by each subset of images specifically includes: For image subsets with insufficient structural clarity, an enhancement operation chain dominated by super-resolution reconstruction and edge sharpening is designed. The operation chain increases image resolution through a deep learning upsampling network and uses an adaptive threshold edge detection and enhancement algorithm to highlight the contours and texture details of power equipment. For a subset of images representing surface anomalies, an enhancement operation chain is designed, which is dominated by local contrast enhancement and color correction. The operation chain estimates the illumination component and enhances the reflection component through Retinex theory, and combines selective saturation adjustment to enhance the visual difference between the defect area and the normal background. For image subsets representing severe environmental interference, a physical model-driven de-degradation enhancement operation chain is designed. For haze images, a dehazing algorithm based on an atmospheric scattering model is used. For rain and snow stripe images, a rain removal algorithm combining time-domain information and frequency-domain filtering is used. For low-illuminance images, a joint algorithm of brightness mapping and noise suppression based on illuminance map estimation is used.

[0011] Preferably, the step of guiding the adaptive adjustment of parameters and selection of operation logic for the image enhancement operation chain based on the structured feature descriptor specifically includes: The structured feature descriptor of a single original image, which is input to the image enhancement operation chain, is used as a control signal to be input to a lightweight policy selection network; The strategy selection network outputs two sets of parameters: one set is a continuous parameter vector used to adjust the intensity of each processing operation in the enhanced operation chain, and the other set is a discrete selection instruction used to determine the specific processing logic in the multi-path processing stage. During the execution of the enhancement operation chain, each algorithm module reads and dynamically configures itself in real time according to the continuous parameter vector and discrete selection instructions, so that the enhancement process is adapted to the specific feature content of the current image, and a synthetic image with the same semantic content as the original image but with enhanced visual representation is generated.

[0012] Preferably, the step of pairing the synthesized image with the corresponding original image, and after authenticity verification and quality filtering, jointly forming the augmented image dataset for model training specifically includes: A realism discriminator based on generative adversarial networks is established, and the discriminator is trained to distinguish between real power platform images and computer-generated fake images; The synthetic image generated by the enhanced operation chain is input into the realism discriminator to filter out synthetic images that are judged to be obviously fake or have unreasonable artifacts; An automatic screening mechanism based on a sharpness evaluation function and information entropy assessment is used to screen the synthetic images that pass the authenticity verification. The synthesized images that have passed verification and filtering will be paired with their original source images and inherit the metadata tags of the original images. Together, they will be included in the augmented image dataset used for model training.

[0013] Compared with the prior art, the advantages of the present invention are: Leveraging a multi-task deep learning model, this system accurately extracts features from three dimensions: power equipment structure, operating status, and scene environment, generating structured feature descriptors to achieve adaptive image clustering. This enables precise differentiation of image samples with different defect types and environmental interference. Customized and diverse enhancement operation chains are tailored for various image subsets, combining feature descriptors to achieve adaptive dynamic adjustment of enhancement parameters and processing logic. This effectively addresses various issues such as image blurring, abnormal exposure, interference from rain, snow, fog, and haze, and unclear defect details, ensuring semantic consistency between the synthesized image and the original image throughout the process. Simultaneously, a dual quality control mechanism—pre-processing multi-layer quality screening and post-processing AI authenticity verification—eliminates invalid and distorted images. This enriches the diversity of power image datasets while ensuring data authenticity and validity, perfectly adapting to complex power inspection scenarios involving multiple perspectives, time periods, and weather conditions. This effectively improves the training effect and generalization ability of the intelligent power equipment recognition model. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of an AI-based image data enhancement method for power platforms proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction and training of a multi-dimensional feature recognition model for power equipment proposed in this invention. Figure 3 This is a schematic diagram of the adaptive adjustment of parameters and selection of operation logic for the image enhancement operation chain proposed in this invention. Detailed Implementation

[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0016] See Figure 1 As shown, an AI-based image data enhancement method for power platforms includes: A raw image dataset from a power platform scenario is collected, and the raw image dataset is preprocessed and quality-screened to construct a benchmark image set. A multi-dimensional feature recognition model for power equipment is constructed and trained. The trained feature recognition model is used to analyze the benchmark image set, extract the structural features, state features and contextual features of the power equipment in each image, and generate corresponding structured feature descriptors. Based on the structured feature descriptor, the baseline image set is adaptively clustered to form multiple image subsets with similar feature attributes; For each subset of images, a diverse set of image enhancement operation chains are designed to match the common features and defect types represented by those features and defects. The image enhancement operation chains are composed of a series of ordered and parameter-adjustable low-level image processing operations. Within each of the image subsets, the parameters of the image enhancement operation chain are adaptively adjusted and the operation logic is selected based on the structured feature descriptor to generate a synthetic image that has the same semantic content as the original image but with enhanced visual representation. The synthesized images are paired with the corresponding original images, and after authenticity verification and quality filtering, they together form an enhanced image dataset for model training.

[0017] Specifically, data is collected from core objects such as power transmission and transformation equipment, lines, towers, and insulators using fixed monitoring equipment, mobile inspection equipment, and drones. The data collection strategy emphasizes multiple perspectives, multiple time periods, and multiple weather conditions to simulate real operation and maintenance environments. Keyframe images are extracted from the video stream at preset time intervals or based on scene motion analysis to form an initial raw image dataset.

[0018] All original images are standardized by converting them to the sRGB color space, and bilinear interpolation and other algorithms are used to uniformly adjust the resolution to a preset standard size. For radial and tangential distortion introduced by the lens, distortion coefficients pre-obtained using a checkerboard calibration method are employed. Geometric correction is performed on the image. The correction transformation can be described by the following model:

[0019]

[0020] Where (x,y) are the normalized pixel coordinates in the original image. These are the corrected coordinates.

[0021] For images that are significantly tilted, the main straight edges in the image are detected by Hough transform, the angle θ between them and the horizontal line is calculated, and the image is rotated by an angle −θ around the center of the image so that the main electrical structures are perpendicular to the horizon.

[0022] Even after standardization, images still undergo rigorous quality screening. Phase one: Image sharpness is primarily characterized by the sharpness of edges and textures, which can be quantified through gradient analysis. This involves converting color images to grayscale. For each pixel (i,j) in the grayscale image, calculate its gradient in the x and y directions. and Then the gradient magnitude of that pixel. for:

[0023] Sharpness score of the whole image It is obtained by calculating the statistical characteristics of the gradient magnitude map, and the mean of the squared gradient magnitudes is used as the evaluation function, which is sensitive to strong edge responses:

[0024] Where N is the total number of pixels in the image. A high score indicates that the image as a whole contains a large number of sharp edges.

[0025] By statistically analyzing a large number of known sharp and blurry images, an empirical threshold is established. If the sharpness score is less than the threshold, the image is judged as "severely blurry" and is discarded.

[0026] Phase Two: Information Entropy measures the information richness of an image region and is used to detect missing content or large invalid areas. The image is divided into K×L regular rectangular grids. For each grid... Statistically analyze the internal grayscale value distribution and calculate the information entropy of its grayscale distribution. :

[0027] in, For grayscale value m in the grid The probability of occurrence in a region. A high entropy value indicates that the region has complex grayscale variations and is rich in information; a low entropy value indicates that the region is uniform or monotonous.

[0028] Based on the prior location of the power equipment, a set of several key grid regions where the equipment may appear is identified. The average entropy value of these key regions is calculated. A threshold is set; if the average entropy value of the key regions is less than the threshold, it indicates that the details of the key equipment area are severely insufficient or information is lost, and the image is deemed invalid and discarded.

[0029] Phase 3: Calculate the global grayscale histogram of the image. If a large number of pixels in the histogram are concentrated at the high-brightness or low-brightness ends, and the region information entropy is extremely low, the image is determined to be either overexposed or underexposed. For images that pass the first two phases, a lightweight semantic segmentation model is run to quickly segment the foreground region of the power equipment. The average information entropy of the pixels in the foreground region is calculated and compared with the average entropy of the entire image. If the average information entropy of the pixels in the foreground region is significantly lower than the average entropy of the entire image, and the foreground region is detected as a large, connected, low-texture block, the critical equipment is determined to be largely occluded and is removed.

[0030] All images retained through the complete screening process described above constitute the final baseline image set. Simultaneously, metadata tags are automatically associated with and recorded for each image, including: acquisition timestamp, device type inferred from image content or source information, and weather conditions derived from acquisition logs or image analysis.

[0031] See Figure 2 As shown, the multi-dimensional feature recognition model for power equipment adopts a deep convolutional neural network architecture with multi-task learning. The front end of the model is a deep convolutional neural network backbone, responsible for extracting multi-level, robust, and general visual features from the input image. The encoder outputs a shared high-level semantic feature map. This contains the contours, textures, and contextual information of all important targets in the image. In shared features... Based on this, the network parallelism derives three independent feature parsing branches, each of which performs refined modeling for a specific cognitive dimension.

[0032] The structural feature branch aims to accurately analyze the geometric structure and inter-component relationships of power equipment. This branch is based on an instance segmentation and keypoint detection architecture, utilizing a region proposal network to... The system locates potential device component regions. For each region, it identifies the component category and its pixel-level segmentation mask, performing component recognition and mask prediction. By predicting the precise coordinates of the component outlines and predefined key connection point coordinates, it detects outlines and key points. Based on the detected components and key points, it automatically constructs a spatial relationship graph G=(V,E), where nodes V represent components and edges E represent connections. The branch outputs a structural integrity vector, which is fused from the component category vector, outline shape descriptor, key point coordinate set, and node and edge features of graph G. Its dimension represents the physical integrity and assembly standardization of the device.

[0033] The state feature branch focuses on identifying abnormal physical states on the device surface, employing a variant of a semantic segmentation network and attaching multiple attribute classification heads. Its input is... After upsampling and feature fusion, a state probability map with the same resolution as the input image is output. State categories include: normal, rust, oil stains, heat, cracks, etc. To enhance sensitivity to subtle defects, this branch integrates channel attention or spatial attention modules, allowing the network to focus on local regions where anomalies may occur. A state index vector is output. This vector is obtained through global pooling and statistical aggregation of the state probability map. It includes the confidence score for each type of anomalous state, as well as quantitative indicators such as pixel area ratio and spatial distribution compactness. A schematic pooling process can be represented as a mask for a specific state k. Global feature aggregation:

[0034] in, For the characteristic transformation function, This indicates element-wise multiplication, and Pooling is a global average pooling operation. Ultimately... By all It was pieced together.

[0035] The environmental features branch is responsible for quantifying external environmental interference factors during image capture, providing context for evaluating image quality and understanding the scene. It includes a global pooling layer and several fully connected layers. It directly accesses shared features. Global context encoding is performed. The output layer consists of multiple parallel sub-output layers, each responsible for predicting a specific environmental factor, such as lighting conditions, fog concentration, rain / snow interference intensity, and background complexity. The output environmental factor vector is a low-dimensional but information-dense vector, directly composed of the predicted values ​​of each environmental factor.

[0036] The final structured feature descriptor output by the model is a high-level fusion of the outputs from the three branches mentioned above. This is typically achieved through a simple concatenation operation or non-linear fusion via a small fully connected network to capture the interactions between features. The model is trained using a large-scale set of labeled power equipment images. The training data must include: structural annotations: component instance segmentation masks, keypoint coordinates; state annotations: pixel-level state category labels; and environment annotations: image-level environmental factor labels.

[0037] The features of each image sample are represented by a structured feature descriptor, which is a high-dimensional vector composed of three parts: structural features, state features, and environmental features, denoted as . To measure the similarity between images, a comprehensive multidimensional feature distance metric is defined. A weighted adaptive distance metric is designed. This is used to calculate the distance between image i and image j:

[0038] in, , , These are independent distance functions for the structural, state, and environmental feature subspaces, respectively. , , These are dynamically adjusted weighting coefficients that reflect the importance and dispersion of different feature dimensions in the current dataset distribution.

[0039] Specifically, since structural feature vectors typically contain category labels and geometric attributes, their distance calculation combines categorical differences and continuous attribute differences. A larger base distance is assigned to different device types; if the types are the same, the Euclidean distance between their contour shape descriptors is calculated. State vectors contain severity indicators for various defects. Their distance can use standardized Manhattan or Euclidean distances, emphasizing the contribution of differences in frequently occurring defects. Environment vectors contain factors such as lighting and weather; their distance is typically calculated using Euclidean distance, but factors that significantly impact image quality, such as fog, rain, and snow, are given higher weights.

[0040] After calculating the feature distance matrix between all image pairs, a density-based spatial clustering algorithm is used for automatic clustering. The neighborhood radius ϵ and the minimum number of points MinPts are defined. For any image sample point p: the ϵ-neighborhood is a spherical region centered at p, with radius ϵ and feature distance d as the metric. If the ϵ-neighborhood of point p contains at least MinPts sample points, then p is a core point. Starting from a core point, all points reachable by its density form a cluster.

[0041] The k-distance graph method is used to adaptively determine parameters, calculating the distance from each point to its MinPts nearest neighbor and sorting them in ascending order. The distance value corresponding to the inflection point in the distance curve is usually taken as the optimal ϵ. Starting from any unvisited core point, all points density-reachable from that point are found, forming a cluster. This process is repeated until all core points have been visited. Points not assigned to any cluster are marked as noise. Each set of maximum density connected samples found by the DBSCAN algorithm that contains at least one core point constitutes an image subset.

[0042] For each cluster Generate a core feature overview and statistical clusters. Structural feature vectors of all images The device types marked in the diagram are selected, with the most frequent type being chosen as the dominant device type for that cluster. :

[0043] in, It is an indicator function. It is the device type label for sample i.

[0044] Analysis cluster The state feature vectors of all images within the cluster. For each defect state, calculate the average confidence or average severity score for that state within the cluster. Select the one or more of the highest scores as the most common one or more defect states in that cluster.

[0045] By analyzing environmental feature vectors, the mean or mode within clusters is calculated for key environmental factors such as light, fog, rain, and snow. For example, if the cluster mean is highest for fog concentration, then foggy weather is the most representative type of environmental disturbance for that cluster.

[0046] For image subsets where device structure is blurred due to long shooting distance, lens resolution limitations, or slight motion blur, an enhancement operation chain is designed to improve high-frequency details and edge sharpness. Lightweight generative adversarial networks or residual dense networks are employed for image super-resolution reconstruction. This network model... Blurry image with low resolution Input: High-resolution image Its objective function includes not only pixel-level mean squared error loss. More importantly, it introduces perceptual loss. and combat losses To ensure that the reconstructed textures conform to the characteristics of natural images:

[0047] in, It is a reference high-resolution image. It is the feature extraction layer of a pre-trained VGG network, used to compute the differences in the high-level feature space.

[0048] For the reconstructed image Edge enhancement is performed. An improved Canny edge detector with high and low thresholds is used. , The gradient magnitude is adaptively calculated based on the statistical distribution of local gradient values ​​in the image, rather than a globally fixed value, to accommodate contrast differences in different regions of the power equipment image.

[0049] in ,and These are the mean and standard deviation of the gradient magnitudes of the entire image. and This is an adjustment coefficient. After edge detection, enhancement is performed using an unsharpening masking technique:

[0050] in, yes Gaussian blur version, The edge enhancement intensity can be adjusted based on the average blur level of the subset.

[0051] For image subsets whose main defects are surface color and texture abnormalities such as rust, oil stains, and dirt, an enhancement operation chain is designed with the separation of illumination and reflection components and local color correction as its core. The multi-scale Retinex algorithm is used to decompose image I(x,y) into illumination component L(x,y) and reflection component R(x,y), with the following model:

[0052] Illumination components are estimated using logarithmic transformation and wrap function: , where F(x,y) is a Gaussian wrapping function. The reflection component passes through The reflection component is considered to contain the object's essential color and texture information, thus mitigating the effects of uneven lighting.

[0053] Adaptive contrast-limited histogram equalization is applied to the reflection component R to enhance the contrast between rusted and oily areas and normal metal surfaces. Simultaneously, selective saturation adjustment is performed based on the color characteristics of specific defects. For example, for rusted areas, the saturation can be non-linearly increased to make the color difference between them and the background more pronounced.

[0054] For image subsets severely affected by environmental factors such as fog, rain, snow, or low illumination, an enhancement operation chain based on the inverse transform of a physical degradation model is designed. For fog / haze interference, the operation chain, based on an atmospheric scattering physics model, estimates global atmospheric light values ​​and scene transmittance maps to reconstruct a clear, fog-free scene from the degraded image. For rain / snow stripe interference, the operation chain utilizes the directional and periodic frequency domain characteristics of these stripes, designing an adaptive filter in the transform domain to remove the corresponding stripe frequency components. Then, in the spatial domain, it combines temporal information for motion compensation and detail restoration to eliminate dynamic stripes and restore the background. For low illumination interference, the operation chain focuses on separating and reconstructing reasonable illumination components from low-brightness images. An adaptive, nonlinear brightness mapping function significantly improves the overall image brightness and contrast, while multi-scale detail fusion technology and edge-preserving filtering are employed to suppress noise amplified by the brightness increase.

[0055] See Figure 3 As shown, a lightweight policy selection network is constructed. The network is a multilayer perceptron, whose input is the feature descriptor of the image and whose output is two types of parameters controlling the augmentation operation chain.

[0056] Strategy Selection Network The forward propagation process can be described as follows:

[0057] Where D is the structured feature descriptor of the input image, W is the trainable weight parameter of the policy selection network, θ is the output continuous parameter vector, which is used to adjust the operation intensity of each processing module in the enhancement operation chain, and z is the output discrete selection instruction vector, which is represented in the form of one-hot encoding or category label, and is used to select specific algorithm branches or logic in the multi-path processing stage of the operation chain.

[0058] network The architecture is designed as an encoder-decoder, with several fully connected layers at the front end encoding the feature descriptors into intermediate semantic vectors; the back end consists of two parallel output heads, one of which outputs continuous parameters θ through a linear layer and a sigmoid activation function, and the other of which outputs discrete selection z through a linear layer and a softmax function.

[0059] The image enhancement operation chain O is designed to consist of L configurable processing modules connected sequentially or in branches. Each module receives specific control parameters from the policy selection network.

[0060] For an input image, the enhancement process can be described as follows: the global control vectors θ and z are parsed into local configuration parameters for each module according to a predefined mapping rule. and selection instructions Each processing module, based on the data it receives... It dynamically adjusts its internal behavior. This manifests in two forms: Parameterized adjustment: The module's internal algorithm parameters are adjusted according to... Real-time changes. For example, in the edge sharpening module, its enhancement intensity factor is directly changed by... One of the components is set. In the dehazing module based on the atmospheric scattering model, the transmittance adjustment factor β can be determined by... Control. Logical path selection: For modules that offer multiple implementation methods, discrete instructions. Used to activate the corresponding processing branch.

[0061] To enable the policy selection network to output effective control parameters, end-to-end optimization is required. This is achieved through a differentiable augmentation operation chain simulation environment. The training objective is to minimize the difference between the augmented image and a high-quality reference target.

[0062] Define a loss function L, which typically includes pixel-level fidelity loss and perceptual loss:

[0063] in, It is the feature extraction layer of a pre-trained VGG network. The sum represents the weight coefficients. Since the entire process is differentiable, the policy selection network can be updated simultaneously via gradient backpropagation. The parameters W, and the parameters of the learnable modules in the enhanced operation chain.

[0064] A dedicated realism discriminator is constructed, which is a deep convolutional neural network. Its core task is to determine whether the input image comes from a real power platform scene or is a synthetic image generated by an enhancement algorithm. It adopts the PatchGAN architecture, which does not output a single real / sorcelain probability for the entire image, but instead outputs a spatial probability map. Each pixel in the map represents the probability that a corresponding local region of the input image is real. The discriminator is optimized through adversarial training, aiming to maximize its ability to distinguish between real and synthetic images. Its loss function... Wasserstein loss with gradient penalty or least squares loss are typically used to improve training stability. Let's take least squares loss as an example: ; in, It is the distribution of real images. It enhances the distribution of the synthetic image generated by the operation chain.

[0065] In the actual filtering phase, the synthesized image to be tested is input into a pre-trained discriminator. The discriminator outputs a spatial probability map. The global average of this probability map is calculated as the realism score of the synthesized image, with a strict empirical threshold set. If the realism score is less than the threshold, the synthesized image is judged to be obviously fake overall or contains large areas of unreasonable regions, and is filtered out. At the same time, the probability map is checked for continuous low probability value regions, which usually correspond to unreasonable artifacts. Such images should be rejected even if the global score meets the threshold.

[0066] Even after the synthetic images pass the authenticity verification, it is still necessary to ensure that they possess sufficient visual information and sharpness to avoid introducing blurry or information-poor invalid samples. An automatic screening mechanism based on a sharpness evaluation function and information entropy assessment is reused to perform a second quality check on the synthetic images.

[0067] The synthesized images, filtered through the two stages described above, establish a bidirectional index relationship between each qualified synthesized image and its corresponding original image, forming data pairs. Each synthesized image fully inherits all metadata tags from its source original image, including acquisition time, device type, weather conditions, and structured feature descriptors obtained after feature recognition model analysis. All paired data, along with original images from the baseline image set that were not covered by the enhancement operation or were already of high quality, are integrated to construct the final enhanced image dataset.

[0068] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0069] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing image data of a power platform based on AI recognition, characterized in that, include: A raw image dataset from a power platform scenario is collected, and the raw image dataset is preprocessed and quality-screened to construct a benchmark image set. A multi-dimensional feature recognition model for power equipment is constructed and trained. The trained feature recognition model is used to analyze the benchmark image set, extract the structural features, state features and contextual features of the power equipment in each image, and generate corresponding structured feature descriptors. Based on the structured feature descriptor, the baseline image set is adaptively clustered to form multiple image subsets with similar feature attributes; For each subset of images, a diverse set of image enhancement operation chains are designed to match the common features and defect types represented by those features and defects. The image enhancement operation chains are composed of a series of ordered and parameter-adjustable low-level image processing operations. Within each of the image subsets, the parameters of the image enhancement operation chain are adaptively adjusted and the operation logic is selected based on the structured feature descriptor to generate a synthetic image that has the same semantic content as the original image but with enhanced visual representation. The synthesized images are paired with the corresponding original images, and after authenticity verification and quality filtering, they together form an enhanced image dataset for model training.

2. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The preprocessing and quality screening of the original image dataset to construct a benchmark image set specifically includes: By using fixed monitoring equipment, mobile inspection equipment or drones deployed at the power platform site, original images and video streams from multiple perspectives, time periods and weather conditions are collected, including power transmission and transformation equipment, lines, towers and insulators, and images are extracted from the video stream at preset time intervals. The acquired raw images are standardized, including: uniformly converting them to a preset color space and pixel resolution, performing optical distortion correction, and aligning the image horizon. An automatic filtering mechanism based on a sharpness evaluation function and information entropy assessment is adopted to remove invalid images that are severely blurry, overexposed, underexposed, or where key power equipment is largely obscured. Images with acceptable image quality are retained to form the reference image set, and each image in the reference image set is labeled with metadata tags indicating the acquisition time, equipment type, and weather conditions.

3. The method for enhancing power platform image data based on AI recognition according to claim 2, characterized in that, The automatic filtering mechanism based on the clarity evaluation function and information entropy assessment specifically includes: The input image is converted into a grayscale image, and the grayscale gradient magnitude and direction of each pixel in the image within its neighborhood are calculated to form a gradient map. Based on the gradient map, a comprehensive sharpness score is obtained by calculating the statistical characteristics of the gradient magnitude. By setting an empirical threshold, images with scores below this threshold are marked as severely blurry, classified as invalid images, and discarded. The image is divided into multiple regular grids, and the information entropy of the gray value distribution is calculated for each grid region. Based on the prior location of the power equipment, we focus on the grid information entropy of the area where the equipment is located. If the information entropy of the area is lower than the set threshold, the image details are severely insufficient, and the image is determined to be invalid and discarded. Analyze the global grayscale histogram of the image. If a large number of pixels in the histogram are concentrated at the high brightness end or the low brightness end, and the regional information entropy is extremely low, then it is determined to be an overexposed or underexposed image. Based on the completion of clarity and information entropy screening, the remaining images are segmented into the foreground region of power equipment. If the detected main equipment area is covered, it is determined that the key equipment is largely occluded and the image is removed. The selected images are then compiled into a baseline image set.

4. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The construction and training of the multi-dimensional feature recognition model for power equipment specifically includes: A deep convolutional neural network for multi-task learning is constructed as the backbone network of the feature recognition model. This backbone network shares the underlying feature extractor and derives three feature parsing branches: structural feature branch, state feature branch, and environment feature branch. The structural feature branches are used to identify the components, outlines, connection points and their spatial topological relationships of specific power equipment in the positioning image, and the structural integrity vector of the equipment is output. The physical state of the equipment surface is analyzed through the state feature branch, including but not limited to the detection and segmentation of rusted areas, oil stains, abnormal heat areas, cracks and damage, and the output of the equipment's state index vector. The overall contextual information of the image is analyzed through the environmental feature branch, including lighting conditions, fog concentration, rain and snow interference, background complexity, and distribution of occlusions, and an environmental factor vector is output. The feature recognition model is trained end-to-end using labeled power equipment image samples until the model stably and accurately outputs the structured feature descriptor, which is a high-dimensional tensor that fuses structural feature vectors, state feature vectors, and environmental feature vectors.

5. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The step of adaptively clustering the baseline image set based on the structured feature descriptor to form multiple image subsets with similar feature attributes specifically includes: Calculate the multidimensional feature distance between the structured feature descriptors of all images in the reference image set; A density-based clustering algorithm is used to automatically aggregate images with similar visual and semantic features into the same cluster based on the feature distance, and each cluster forms a subset of the images. For each of the image subsets, a core feature overview is summarized, which includes the dominant device type, the most common one or more defect states, and the most representative environmental interference type within the subset.

6. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The design of diverse image enhancement operation chains that match the common features and defect types represented by each subset of images specifically includes: For image subsets with insufficient structural clarity, an enhancement operation chain dominated by super-resolution reconstruction and edge sharpening is designed. The operation chain increases image resolution through a deep learning upsampling network and uses an adaptive threshold edge detection and enhancement algorithm to highlight the contours and texture details of power equipment. For a subset of images representing surface anomalies, an enhancement operation chain is designed, which is dominated by local contrast enhancement and color correction. The operation chain estimates the illumination component and enhances the reflection component through Retinex theory, and combines selective saturation adjustment to enhance the visual difference between the defect area and the normal background. For image subsets representing severe environmental interference, a physical model-driven de-degradation enhancement operation chain is designed. For haze images, a dehazing algorithm based on an atmospheric scattering model is used. For rain and snow stripe images, a rain removal algorithm combining time-domain information and frequency-domain filtering is used. For low-illuminance images, a joint algorithm of brightness mapping and noise suppression based on illuminance map estimation is used.

7. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The adaptive adjustment of parameters and selection of operation logic for the image enhancement operation chain guided by the structured feature descriptor specifically includes: The structured feature descriptor of a single original image, which is input to the image enhancement operation chain, is used as a control signal to be input to a lightweight policy selection network; The strategy selection network outputs two sets of parameters: one set is a continuous parameter vector used to adjust the intensity of each processing operation in the enhanced operation chain, and the other set is a discrete selection instruction used to determine the specific processing logic in the multi-path processing stage. During the execution of the enhancement operation chain, each algorithm module reads and dynamically configures itself in real time according to the continuous parameter vector and discrete selection instructions, so that the enhancement process is adapted to the specific feature content of the current image, and a synthetic image with the same semantic content as the original image but with enhanced visual representation is generated.

8. The method for enhancing power platform image data based on AI recognition according to claim 1, characterized in that, The process of pairing the synthesized image with the corresponding original image, and then performing authenticity verification and quality filtering to jointly form the augmented image dataset for model training specifically includes: A realism discriminator based on generative adversarial networks is established, and the discriminator is trained to distinguish between real power platform images and computer-generated fake images; The synthetic image generated by the enhanced operation chain is input into the realism discriminator to filter out synthetic images that are judged to be obviously fake or have unreasonable artifacts; An automatic screening mechanism based on a sharpness evaluation function and information entropy assessment is used to screen the synthetic images that pass the authenticity verification. The synthesized images that pass the verification and filtering process are paired with their original source images and inherit the metadata tags of the original images. Together, they are entered into the augmented image dataset used for model training.

Images