Power fitting corrosion detection method and system fusing open vocabulary and structure perception

Abstract

This invention discloses a method and system for detecting corrosion of power fittings that integrates open vocabulary and structure awareness. The method employs a two-stage detection architecture: In the first stage, a CLIP-LoRA multimodal model, fine-tuned with power domain semantic guidance, is introduced into the Grounding DINO backbone network to enhance the contextual semantic understanding capability of open vocabulary fitting detection; multi-scale features are extracted in parallel through standard convolutions and dilated convolutions with different dilation rates, combined with spatial and channel attention mechanisms to improve the localization accuracy of fittings of unknown categories; In the second stage, a windmill-shaped convolution is introduced into the YOLOv11 backbone network to expand the receptive field while maintaining constant parameters, and a scale-aware pyramid pooling module is designed to integrate multi-scale corrosion features. This invention effectively improves the accuracy and generalization ability of power fitting corrosion detection.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology for power transmission line fittings, specifically to a method and system for detecting corrosion of power fittings that integrates open vocabulary and structure perception. Background Technology

[0002] Transmission line fittings, as a crucial component of power infrastructure, serve to protect, secure, support, and maintain the stability of the line structure. During long-term operation, these fittings are susceptible to corrosion due to environmental and operational conditions. If this corrosion is not addressed promptly, it can lead to abnormal line operation and even serious power accidents and economic losses. Therefore, regular inspection and maintenance of transmission lines are essential measures to prevent major faults and ensure reliable power supply from the grid.

[0003] In recent years, computer vision and artificial intelligence technologies have made significant progress in the field of power transmission line defect detection. Researchers have pre-trained object detection models on large-scale general datasets such as ImageNet and COCO, and then fine-tuned them on domain-specific datasets containing images of power transmission line defects. This approach achieves efficient defect detection and significantly improves the efficiency of power line inspection. However, general object detection models often lack the ability to model specific domain knowledge related to the power industry, resulting in poor performance when directly applied to power transmission line defect detection tasks. To overcome this limitation, current research mainly focuses on two directions: first, enhancing the model's sensitivity to defect features by optimizing the network architecture; and second, improving the model's generalization ability in small sample sizes or complex scenarios by refining training strategies. In addition, some studies have explored integrating textual descriptions related to power transmission line defects to construct domain knowledge graphs, aiming to improve detection accuracy and robustness. In broader vision tasks, Zang et al. introduced Vision-Language Models (VLMs) into the object detection framework, achieving a breakthrough in open-world object detection. In the field of medical image analysis, VLMs have also been applied to lesion detection and segmentation tasks, significantly improving the model's ability to recognize complex pathological features. These studies demonstrate that the strong alignment between image and text features provided by VLMs can significantly improve the model's performance in multimodal tasks.

[0004] With the continuous development of vision-language models, such as contrastive language-image pre-training (CLIP) and foundational language and vision alignment (FLAV), methods that can effectively align image and text descriptions have gradually emerged. Related research has proposed fine-tuning language-image models, such as the low-rank fine-tuning model for vision-language (CLIP-LoRA). The paradigm of aligning text and visual features in vision-language models has demonstrated its effectiveness in improving the accuracy of downstream tasks. Among these methods, GLIP, GLIPv2, OV-DETR, DetCLIP, YOLO-World, YOLO-UniOW, and Grounding DINO have all shown excellent object detection performance. For the detection of corrosion of power fittings, existing technologies have recorded the integration of edge perception and statistical texture knowledge, using the Res-UNet segmentation model to perform edge perception on the fittings of transmission lines, thereby achieving accurate location of defects. In addition, SE (squeeze-excitation)

[26] and SimAM (simple parameter-free attention module) are introduced into the YOLOv7 network to enhance the network's perception of complex environments. In terms of multimodal power target defect detection, existing technologies have also designed a detection model based on visual language and high-frequency features. This model is effective in Deformable The DETR framework employs a multi-stage training strategy. In the first stage, to address the lack of domain-specific knowledge in the general visual language model's base language and visual alignment (FLAVA), FLAVA is fine-tuned to learn textual and visual features relevant to the power industry. Furthermore, structural features of inspection images are utilized, and Fast Fourier Transform (FFT) is applied to extract high-frequency edge features, suppress background noise, and provide spatial priors. An FFT-based loss function (FFTLoss) is introduced to further ensure model convergence on the target region. Overall, the above methods have achieved considerable progress in power fitting defect detection, but several challenges and problems remain to be solved. First, power fittings exhibit drastic scale variations, diverse types, and complex scenarios, making it difficult to design effective network modules to handle the detection of multiple types of fittings in complex scenarios. Second, the irregular morphology of rusted areas on fittings means that traditional feature extraction backbone networks using conventional convolutional designs fail to address the specific difficulties in rust detection. These problems pose serious challenges to substation defect detection, necessitating further research into solutions. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method and system for detecting corrosion of power fittings that integrates open vocabulary and structure awareness, in order to address the above-mentioned problems of the prior art. This invention aims to achieve more accurate detection of corrosion on the surface of fittings, establish structured target associations, and enhance the multi-scale structure awareness capability of the network to meet the challenges of multi-scale and multi-class detection.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting corrosion of power fittings that integrates open vocabulary and structure awareness includes the following steps: S1, fine-tuning a backbone network using an image-text dataset from the power industry, wherein the backbone network includes an image encoder and a text encoder; S2, embedding the fine-tuned backbone network into a power fitting detection network and a fitting corrosion detection network, wherein the power fitting detection network is used to locate the position of power fittings from the input image based on the input text features, and the fitting corrosion detection network is used to detect corrosion from the power fitting image. The fitting detection network is trained using an open vocabulary fine-tuning dataset, and closed sets are utilized... The rust dataset completes the training of the rust detection network. The open vocabulary fine-tuning dataset includes the object's bounding box and the corresponding category's text description. The closed set rust dataset only contains rust images and rust category information. S3, the input image and its corresponding text features are used to detect power fitting images using the trained fitting detection network. The text features are tensors obtained by feature extraction of text through a trained text encoder. The text features have a one-to-one correspondence with the category of the object in the input image. The detected power fitting images are then used to complete rust detection using the trained fitting rust detection network.

[0007] Optionally, the backbone network in step S1 is the Grounding DINO backbone network, and the fine-tuning includes using an image-text dataset and a contrastive loss function from the power field, and using LoRA technology to fine-tune the Grounding DINO backbone network.

[0008] Optionally, the power fitting detection network includes a dual-branch multi-scale structure perception module (DMSM), a feature enhancement module, a language-guided query selection module, and a cross-modal decoder. The DMSM is used to perform structure perception on the image features extracted by the image encoder. The feature enhancement module is used to enhance the image features after structure perception and the text features extracted by the text encoder. The language-guided query selection module is used to perform language-guided queries on the enhanced image features and text features to obtain query features. The cross-modal decoder is used to perform cross-modal decoding based on the enhanced image features, text features, and query features to obtain the location box and confidence score of the power fitting.

[0009] Optionally, the dual-branch multi-scale structure perception module uses a dual-branch structure consisting of two branches to extract multi-scale features, and then passes the multi-scale features sequentially through a spatial attention module and a channel attention module to obtain structure-perceived image features. Of the two branches, the first branch includes two cascaded 3×3 combined convolutional modules (CBR3), each composed of a 3×3 convolutional module, a batch normalization layer (BN), and a ReLU activation unit. The second branch includes two cascaded combined dilated convolutional modules (DCBR), each composed of dilated convolution, a batch normalization layer (BN), and a ReLU activation unit, with different dilation rates in the two DCBR modules. The functional expression for extracting multi-scale features using the dual-branch structure is as follows: ; in, For multi-scale features, It is a 1×1 combined convolutional module, which consists of a 3×3 convolutional module CBR, a batch normalization layer BN, and a ReLU activation unit; It is a 3×3 combined convolution module. As input features, For the combined dilated convolutional module DCBR with a dilation rate of 2, It is a combined dilated convolutional module DCBR with a dilation rate of 4.

[0010] Optionally, the feature enhancement module is composed of multiple stacked enhancement layers with the same structure. The enhancement layers perform feature enhancement on the image features after structure perception and the text features extracted by the text encoder, including: first, performing intra-modal enhancement on the image features and text features respectively through parallel self-attention; then, performing inter-modal fusion on the intra-modal enhanced image features and text features through cross-attention to exchange semantic information between modalities; and finally, outputting the inter-modal fused image features and text features after nonlinear transformation by a feedforward network.

[0011] Optionally, the language-guided query selection module is used to perform language-guided queries on the enhanced image features and text features to obtain query features, including: calculating the dot product similarity between the image features and text features respectively to obtain a similarity matrix composed of dot product similarity; for each image feature position in the image features, taking the maximum value of the similarity with all text features in the similarity matrix to form a score vector representing the overall relevance between the image feature position and the text features; finally, selecting the top k image feature positions with the highest scores from the score vector and extracting the corresponding image features based on these position indices as the obtained query features.

[0012] Optionally, the cross-modal decoder includes multiple stacked decoding layers and a detection head. The decoding layers perform cross-modal decoding based on the enhanced image features, text features, and query features to obtain images of power fittings. This includes: extracting the target information of each input image feature, text feature, and query feature through self-attention layers to obtain enhanced image features, text features, and query features; enabling intermodal interaction between the query features and text features through cross-attention layers to fuse language semantics for target localization; enabling intermodal interaction between the query features and enhanced image features after intermodal interaction through cross-attention layers to capture visual details; outputting the query features after completing two intermodal interactions through a feedforward network for feature transformation, and then using the detection head to regress and classify the features output by the last decoding layer to generate bounding boxes and category prediction results.

[0013] Optionally, in step S2, when training the hardware detection network and the hardware corrosion detection network using the open vocabulary fine-tuning dataset, the loss function used includes scale-adaptive loss, the expression for which the scale-adaptive loss is calculated is: ; ; ; ; in, For scale-adaptive loss, For scale adaptation, It is a natural constant. To predict the area of ​​the bounding box, For constant hyperparameters, The intersection-union ratio (IUU) of the predicted bounding box and the target bounding box. For constant hyperparameters Normalized Wasserstein distance, For constant hyperparameters Normalized Wasserstein distance, For prediction boxes and target box The Wasserstein distance between Gaussian distributions For prediction boxes and target box The L2 norm.

[0014] Optionally, the metal fitting corrosion detection network is an improved YOLOv11 model. The improved YOLOv11 model consists of a target detection backbone network, a neck network, and a head network. The target detection backbone network comprises two cascaded convolutional layers, four feature extraction units, a fast spatial pyramid pooling module (SPPF), and a scale-aware context pyramid pooling module (SCPP). The feature extraction unit consists of an improved C3K2 module and a convolutional layer. The improved C3K2 module is obtained by replacing the 3×3 convolutional module in the bottleneck layer (Bottleneck) of the original C3K2 module with a windmill-shaped convolutional module (PConv). The scale-aware context pyramid pooling module (SCPP) includes three parallel branches, where the first branch is used to process the input feature F... input The first output feature is obtained through a 1×1 convolution, and the second branch converts the input feature F... input The second output feature is obtained by using a multi-scale module, and the third branch extracts the input feature F. input Global average pooling is used to obtain weights. The first and second output features are concatenated by the channel dimension and then multiplied by the weights. This allows the output features of the Scale-Aware Context Pyramid Pooling (SCPP) module to be obtained through a 1×1 convolution. The multi-scale module includes three parallel sub-branches, each of which includes a dilated convolution, a compression activation module, and a 1×1 convolution. The dilation rates of the dilated convolutions in the three sub-branches are different. The feature maps output by the 1×1 convolutions of the three sub-branches are summed and then subjected to a 3×3 convolution and a Softmax function to generate pixel-level attention maps of three scales. , and The second output feature is calculated according to the following formula. : ; in, For element-wise multiplication, , and These are the feature maps output by the 1×1 convolution of the three sub-branches.

[0015] Furthermore, the present invention also provides a corrosion detection system for power fittings that integrates open vocabulary and structure awareness, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the corrosion detection method for power fittings that integrates open vocabulary and structure awareness.

[0016] Compared with the prior art, the present invention can mainly achieve the following beneficial effects: (1) The present invention proposes a two-stage power fitting corrosion detection method that integrates open vocabulary and structure perception. It uses open vocabulary multimodal detection as a pre-guide and is supplemented by a structure perception module to achieve faster and more accurate surface corrosion detection of fittings; (2) The present invention combines the power fitting corrosion detection method that integrates open vocabulary and structure perception with the backbone network fine-tuned by power multimodal data and the dual-branch multiscale structure perception module to effectively establish structured target association and enhance the network's multiscale structure perception capability to meet the challenges of multiscale and multi-class detection; (3) The present invention integrates the power fitting corrosion detection method that integrates open vocabulary and structure perception. In the second stage, it designs a mechanism that integrates windmill-shaped convolution and multiscale feature fusion to improve the perception and positioning accuracy of irregular corrosion areas under the premise of high parameter efficiency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the network structure in the first stage of an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of the network structure in the second stage of an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the network structure of the dual-branch multi-scale structural sensing module DMSM in an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of the network structure in the third stage of an embodiment of the present invention.

[0022] Figure 6 This is a schematic diagram of the network structure of the windmill-shaped convolutional PConv in an embodiment of the present invention.

[0023] Figure 7 This is a schematic diagram of the network structure of the Scale-Aware Context Pyramid Pooling Module (SCPP) in an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1As shown, the power fitting corrosion detection method integrating open vocabulary and structure awareness in this embodiment includes the following steps: S1, fine-tuning the backbone network using an image-text dataset from the power industry, wherein the backbone network includes an image encoder and a text encoder; S2, embedding the fine-tuned backbone network into a power fitting detection network and a fitting corrosion detection network, wherein the power fitting detection network is used to locate the position of power fittings from the input image based on the input text features, and the fitting corrosion detection network is used to detect corrosion from the power fitting image. The fitting detection network is trained using an open vocabulary fine-tuning dataset, and the method utilizes... The closed-set corrosion dataset is used to train the corrosion detection network. The open-vocabulary fine-tuning dataset includes the object's bounding box and the corresponding category's text description. The closed-set corrosion dataset only contains corrosion images and corrosion category information. S3, the input image and its corresponding text features are used to detect power equipment images using the trained equipment detection network. The text features are tensors obtained by feature extraction of text through a trained text encoder. The text features have a one-to-one correspondence with the category of the object in the input image. The detected power equipment images are then used to complete corrosion detection using the trained equipment corrosion detection network.

[0026] This embodiment of the power fitting corrosion detection method, which integrates open vocabulary and structure awareness, can be divided into three parts: steps S1 to S3. For example... Figure 2 As shown, the backbone network in step S1 of this embodiment is the Grounding DINO backbone network (CLIP backbone network). The fine-tuning includes using an image-text dataset and a contrastive loss function from the power industry, and fine-tuning the Grounding DINO backbone network using LoRA technology, referred to as the CLIP-LoRA method. The first stage of step S1 is the power multimodal fine-tuning part, which uses power multimodal image-text pair data to fine-tune the CLIP-LoRA method for power scene, realizing the mining and alignment of power target semantic information. The contrastive loss is the average of the image-to-text loss and the text-to-image loss. Since the image-to-text loss and the text-to-image loss are both known losses, their implementation details will not be described here. The CLIP-LoRA multimodal model fine-tuning method solves, to some extent, the problem of generalizing pre-trained models from the general domain to the power industry. The low-rank adaptation (LoRA) fine-tuning method is added to the classic CLIP model, and the incremental update of the pre-trained weight fine-tuning is modeled as two small matrices. A and B The product of the inputs. x Hidden state h and weight matrix The forward propagation process after applying the LoRA module can be represented as: ; in, , The rank of the matrix is of ( Typically much smaller than { , }), where γ is the scaling factor. Matrix A The values ​​in the matrix are randomly generated through Kaiming initialization, while the matrix... B It is initialized to an all-zero matrix. This means that there are no incremental updates before training begins, so the model's output remains unchanged.

[0027] In the CLIP-LoRA method, a low-rank matrix is ​​applied to the attention matrix based on the Transformer architecture.

[32] This architecture is typically composed of L It consists of stacked modules, each containing a multi-head attention (MHA) module: ; ; in d Scaling factor , , , These are the key, query, value, and output matrix, respectively.

[0028] Standard 3×3 or 5×5 convolutions are primarily used to extract local spatial information, but they struggle to effectively establish semantic associations for structured targets. Outdoor lighting interference and variations in target contrast further exacerbate this problem, posing a significant challenge to the detection of structured multi-target objects like power fittings. Furthermore, the detection of power fittings often utilizes drones, resulting in varying scale distributions from the shooting perspective, rendering single-scale feature extraction methods insufficient to capture multi-scale target changes. To address these challenges, the power fitting detection network in this embodiment includes a dual-branch multi-scale structure perception module (DMSM), a feature enhancement module, a language-guided query selection module, and a cross-modal decoder. The DMSM performs structure perception on image features extracted by the image encoder. The feature enhancement module enhances the structured image features and text features extracted by the text encoder. The language-guided query selection module performs language-guided queries on the enhanced image and text features to obtain query features. The cross-modal decoder performs cross-modal decoding based on the enhanced image, text, and query features to obtain the bounding box and confidence score of the power fitting.

[0029] like Figure 3 As shown, in step S2, when embedding the fine-tuned backbone network into the power fitting detection network and the fitting corrosion detection network, the image encoder and text encoder of the Grounding DINO backbone network are frozen. Then, the power fitting detection network extracts power fitting images from the input image based on the input text features, and the fitting corrosion detection network detects corrosion from the power fitting images. The training of the fitting detection network and the fitting corrosion detection network is completed using an open vocabulary fine-tuning dataset.

[0030] like Figure 4 As shown, the dual-branch multi-scale structure sensing module (DMSM) is added after the image encoder output. The DMSM enhances the receptive field and feature representation by combining multi-scale feature extraction with spatial and channel attention mechanisms. It consists of two core components: dual-branch feature extraction and a dual-attention mechanism. Given an input feature map... This network employs a dual-branch structure to extract multi-scale features. The first branch uses standard convolution operations to maintain local spatial consistency and extract features. The second branch fuses dilated convolutions with multiple dilation rates to expand the receptive field and capture contextual information at different scales. This dual-branch structure can extract context at different scales, enrich feature representations, and ensure that the network maintains a balance between local details and context. Figure 2 As shown, in this embodiment, the dual-branch multi-scale structure perception module (DMSM) uses a dual-branch structure consisting of two branches to extract multi-scale features. These multi-scale features are then sequentially passed through a spatial attention module and a channel attention module to obtain structure-aware image features. The first branch includes two cascaded 3×3 combined convolutional modules (CBR3), each composed of a 3×3 convolutional module, a batch normalization layer (BN), and a ReLU activation unit. The second branch includes two cascaded combined dilated convolutional modules (DCBR), each composed of dilated convolutions, a batch normalization layer (BN), and a ReLU activation unit. The dilation rates of the dilated convolutional modules in the two DCBR modules are different. The functional expression for extracting multi-scale features using the dual-branch structure is as follows: ; in, For multi-scale features, It is a 1×1 combined convolutional module, which consists of a 3×3 convolutional module CBR, a batch normalization layer BN, and a ReLU activation unit; It is a 3×3 combined convolution module. As input features, For the combined dilated convolutional module DCBR with a dilation rate of 2, The DCBR (Dilated Convolutional Layer) module is used, with a dilation rate of 4. After multi-scale feature extraction, spatial and channel attention mechanisms are employed to refine and fuse the features, ensuring that the network focuses on relevant target regions while suppressing background noise. The structure-aware image features obtained by sequentially passing the multi-scale features through the spatial and channel attention modules can be described as follows: ; In the formula, This indicates the output of the DMSM. and Let SAM and CAM represent the spatial attention module and channel attention module, respectively. The spatial attention module (SAM) and channel attention module (CAM) can be represented as follows: ; ; in, Represents the given input features, This represents the sigmoid function. This represents a convolution operation with a kernel size of 7×7. and These represent max pooling and average pooling operations with a step size of 2, respectively. Indicates element-wise multiplication. It is a multi-layer learning machine. By employing a dual-branch feature extraction structure and attention-guided optimization, the dual-branch multi-scale structure perception module (DMSM) achieves robust multi-scale feature representation, which is beneficial for the detection of structured power fittings. Combining dilated convolutions for context enhancement and attention mechanisms for target focusing feature refinement significantly improves the detection performance of fittings in complex power scenarios.

[0031] In this embodiment, the feature enhancement module is composed of multiple enhancement layers with the same structure stacked together. The enhancement layers perform feature enhancement on the image features after structure perception and the text features extracted by the text encoder, including: first, performing intramodal enhancement on the image features and text features respectively through parallel self-attention; then, achieving multimodal enhancement through cross-attention, and performing intermodal fusion on the image features and text features to exchange semantic information between modalities; finally, outputting the intermodal fused image features and text features after nonlinear transformation by a feedforward network.

[0032] In this embodiment, the language-guided query selection module is used to perform language-guided queries on the enhanced image features and text features to obtain query features. The dot product similarity between the image features and text features is calculated to obtain a similarity matrix composed of dot product similarities. For each image feature location in the image features, the maximum similarity with all text features in the similarity matrix is ​​taken to form a score vector representing the overall relevance between the image feature location and the text features. Finally, the top k image feature locations with the highest scores are selected from this score vector, and the corresponding image features are extracted based on these location indices as the obtained query features.

[0033] In this embodiment, the cross-modal decoder includes multiple stacked decoding layers and a detection head. The decoding layer performs cross-modal decoding based on the enhanced image features, text features, and query features to obtain images of power fittings. This includes: extracting the target information of each input image feature, text feature, and query feature through self-attention layers to obtain enhanced image features, text features, and query features; enabling intermodal interaction between the query features and text features through cross-attention layers to fuse language semantics to guide target localization; enabling intermodal interaction between the query features and enhanced image features after intermodal interaction through cross-attention layers to capture visual details; outputting the query features after completing two intermodal interactions through a feedforward network for feature transformation, and using the detection head to locate and classify the features output by the last decoding layer to generate bounding boxes and category prediction results.

[0034] In step S2 of this embodiment, when training the hardware detection network and the hardware corrosion detection network using the open vocabulary fine-tuning dataset, the loss function used includes scale-adaptive loss. The calculation function expression of the scale-adaptive loss is as follows: ; ; ; ; in, This is the scale-adaptive loss (SAFit loss). For scale adaptation, It is a natural constant. To predict the area of ​​the bounding box, For constant hyperparameters, The intersection-union ratio (IUU) of the predicted bounding box and the target bounding box. For constant hyperparameters Normalized Wasserstein distance, For constant hyperparameters Normalized Wasserstein distance, For prediction boxes and target box The Wasserstein distance between Gaussian distributions For prediction boxes and target box The L2 norm. The Normalized Wasserstein Distance (NWD) loss, due to its scale invariance and smoothness against positional bias, has proven highly effective for small target detection. Among them... It is a predictive regression box = and target box = The Wasserstein distance between the Gaussian distributions, with the center point located at... ,width w and height h . K It is a hyperparameter closely related to the data. However, the scale invariance of NWD cannot provide a reasonable evaluation for objects with size changes. Intersection over Union (IoU) is a commonly used and reasonable metric for evaluating the performance of object detection methods. The formula can be defined as: ; in, and Let represent the predicted bounding box and the target bounding box, respectively. Although IoU has advantages such as scale invariance and symmetry, it has low tolerance for IoU perturbations in small objects. Small positional deviations (e.g., a 2-pixel deviation for an 8×8 target) can lead to a significant decrease in IoU. In summary, using IoU-based metrics to evaluate object detection methods has certain limitations. To combine the advantages of IoU and NWD while avoiding their disadvantages, this embodiment employs a scale-adaptive loss, which is highly robust to both large and small objects. Specifically, IoU and NWD are combined using a size-aware sigmoid weighted summation, i.e.: ; The Sigmoid function represents a soft switch, which can quickly switch the size of the prediction box to an appropriate scale. A is the area of ​​the true bounding box, and C is a constant that balances the NWD and IoU metrics in a scale-aware manner. That is, when A = At that time, NWD and IoU contribute equally. Lower AA higher A value (meaning a smaller bounding box) leads to a higher proportion of NWD (Negative Dimensioning Loss), while a higher A value (meaning a smaller bounding box) leads to a higher proportion of IoU (Intersection over Union). In summary, the SAFit loss (Scale Adaptive Loss) has significant advantages for practical applications involving objects of different scales. It is worth noting that by adjusting... C The value of SAFit loss allows for flexible applications to meet various custom requirements.

[0035] like Figure 5 As shown, as an optional implementation, the hardware corrosion detection network in this embodiment is an improved YOLOv11 model. The improved YOLOv11 model consists of a target detection backbone network, a neck network, and a head network. The target detection backbone network comprises two cascaded convolutional layers (2×convolutional layers), four feature extraction units, a fast spatial pyramid pooling module (SPPF), and a scale-aware context pyramid pooling module (SCPP). The feature extraction units consist of an improved C3K2 module and a convolutional layer (Conv). The improved C3K2 module is obtained by replacing the 3×3 convolutional module in the bottleneck layer (Bottleneck) of the original C3K2 module with a windmill-shaped convolutional layer (PConv), which significantly increases the receptive field of the target detection backbone network. The bottleneck layer (Bottleneck) of the improved C3K2 module consists of a 3×3 convolution, a windmill-shaped convolutional layer (PConv), and a residual module. The residual module is used to add the output features of the windmill-shaped convolutional layer (PConv) to the input features of the bottleneck layer (Bottleneck) as the output features of the bottleneck layer (Bottleneck).

[0036] like Figure 6 As shown, the windmill-shaped convolution PConv includes four branches, each branch including a padding module and a 2D convolutional block. The padding modes of the padding modules in the four branches are (1,0,0,3), (0,3,0,1), (0,1,3,0), and (3,0,1,0), respectively. The four values ​​of the padding mode represent the number of padding pixels on the left, right, top, and bottom, respectively. For example, Padding(1,0,0,3) in the first branch represents a padding module with a padding mode of (1,0,0,3). The convolutional kernel sizes of the 2D convolutional blocks in the four branches are 1×3, 3×1, 1×3, and 3×1, respectively. For example... Figure 6The first branch, `Conv(c,(1,3),s,0)`, represents a 2D convolutional block with output channels `c`, a kernel size of 1×3, a stride of `s`, and zero padding. Finally, the output features of the 2D convolutional blocks from the four branches are concatenated along the channel dimension to obtain the final output features. This output feature is then passed through a 2D convolutional block, represented by `Conv(c,(2,1),1,0)`, with output channels `c`, a kernel size of 2×2, a stride of 1, and zero padding, to obtain the final output features of the windmill-shaped convolution `PConv`. The receptive field weights of the windmill-shaped convolution `PConv` decrease outwards, similar to a Gaussian distribution. Furthermore, the smaller the target, the more concentrated its features, highlighting the importance of the central features.

[0037] In metal corrosion detection, high intra-class variability makes it difficult to distinguish object features, easily leading to false detections and missed detections. Therefore, fully utilizing features at different scales is crucial for metal corrosion detection. To address these challenges, dilated convolution aims to enhance feature representation by increasing the receptive field, and dilated convolution with different dilation rates enhances context awareness. However, the drawback of cross-pixel sampling in dilated convolution leads to information reduction, resulting in the ineffective perception and utilization of specific locations determined by the dilation rate. To improve this problem, a module needs to be designed to overcome the drawbacks of dilated convolution while leveraging its large receptive field. The Scale-Aware Context Pyramid Pooling (SCPP) module overcomes the shortcomings of dilated convolution while providing multi-scale awareness. Using the scale-aware module, important features can be adaptively focused on and multi-scale information can be fully utilized, mitigating false detections and missed detections caused by high intra-class variability. Figure 7 As shown, the Scale-Aware Context Pyramid Pooling Module (SCPP) includes three parallel branches, where the first branch is used to process the input features F. input The first output feature is obtained through a 1×1 convolution, and the second branch converts the input feature F... input The second output feature is obtained by using a multi-scale module, and the third branch extracts the input feature F. input Global average pooling is used to obtain weights. The first and second output features are concatenated by the channel dimension and then multiplied by the weights. This allows the output features of the Scale-Aware Context Pyramid Pooling (SCPP) module to be obtained through a 1×1 convolution. The multi-scale module includes three parallel sub-branches, each of which includes a dilated convolution, a compression activation module, and a 1×1 convolution. The dilation rates of the dilated convolutions in the three sub-branches are different. In this embodiment, the dilation rates of the dilated convolutions in the three sub-branches are 1, 3, and 5, respectively. The feature maps output by the 1×1 convolutions of the three sub-branches are summed and then subjected to a 3×3 convolution and a Softmax function to generate pixel-level attention maps of three scales. , and The second output feature is calculated according to the following formula. : ; in, For element-wise multiplication, , and The feature maps output by the 1×1 convolutions of the three sub-branches are presented separately. Finally, the resulting multi-scale feature fusion map is concatenated with the convolutional branches and the global pooling branch to supplement contextual information, thereby obtaining the final feature representation. Through the processing of the scale-aware module, SCPP exhibits excellent multi-scale structure perception capabilities and effectively alleviates the detection errors and missed detection problems of metal fitting corrosion. Figure 7 In the diagram, H, W, and C represent the input features F, respectively. input Height, width, and number of channels.

[0038] To verify the effectiveness of the power fitting corrosion detection method integrating open vocabulary and structure awareness in this embodiment, the dataset used in this embodiment is a collection of fittings and corrosion images specifically collected for power fitting corrosion detection. It is formatted according to the data format required by the method in this embodiment to create a fitting corrosion detection dataset, which includes three sub-datasets: (1) a power multimodal fine-tuning dataset, containing 40,000 image-text pairs related to power inspection scenarios; (2) a power fitting open vocabulary target detection dataset, containing 15,793 power fitting target boxes, with 20 fitting categories, and the category labels are converted into text statements using a script; and (3) a fitting corrosion target detection dataset, containing 1,264 fitting corrosion target boxes. Table 1 shows the detailed category data of the power fitting open vocabulary target detection dataset.

[0039] Table 1: Open Vocabulary Object Detection Dataset for Power Fittings

[0040] This dataset integrates four existing open-source power fitting datasets: InsPLAD, STN PLAD, PTL-AIFurnas, and EPFD. It covers fitting images with different backgrounds, resolutions, shooting angles, and shooting distances to ensure sample diversity and consistency with real-world power fitting scenarios. After screening, 4352 high-quality power fitting images were obtained, and after integration and sorting, 15793 bounding boxes were obtained across 20 classes. This satisfies the requirements for training and validation data for open-vocabulary object detection. Furthermore, in terms of open-vocabulary data partitioning, five categories—right-angle brackets, wire clips, U-bolts, wedge clamps, and tower plates—were treated as unknown categories, while the remaining 15 categories were treated as known categories for training. The experiment used Ubuntu 20.04 as the operating system, an Intel(R) i9-14900K CPU with a clock speed of 2.4GHz, 128GB of RAM, and two NVIDIA GeForce RTX 4090 graphics cards with 48GB of VRAM, using CUDA 11.8 for accelerated computation. The deep learning framework chosen was PyTorch 2.2.0, compatible with CUDA 11.8, with Python version 3.10.19. In the training process of fine-tuning the CLIP-LoRA multimodal model, this embodiment uses 2×10... -4 The learning rate was set to a cosine decay strategy, the batch size was 32, the optimizer was stochastic gradient descent (SGD) with a momentum parameter of 0.93 and a weight decay coefficient of 0.0005, and the training epochs were set to 100. The Grounding DINO method, an open-vocabulary object detection method, was trained with a learning rate of 1×10⁻⁶. -3 The learning rate was set using a cosine learning rate decay strategy, with a batch size of 8, and optimizer parameters were set the same as in the CLIP-LoRA method. Finally, the training parameters for the YOLOv11-based improved metal corrosion detection method were kept at their default values, with a batch size of 16. All training was performed on a dual-GPU setup with a capacity of 48 Gb.

[0041] In this embodiment, three sets of comparative experiments were conducted: a comparative experiment on open-vocabulary hardware detection, a comparative experiment on hardware corrosion detection, and a comparative experiment on two-stage hardware detection. On the open-vocabulary target detection dataset for power hardware, the improved Grounding DINO model, composed of a power hardware detection network and a hardware corrosion detection network, proposed in this embodiment was evaluated and compared with existing models such as GLIP, GLIPv2, OV-DETR, DetCLIP, YOLO-World, YOLO-UniOW, and Grounding DINO. The results are shown in Table 2.

[0042] Table 2: Comparative experimental results of the improved Grounding DINO model

[0043] Table 2 presents the experimental results of the method proposed in this embodiment compared with other comparative methods. The results in Table 2 show that, in the experimental results for known categories, the baseline method Grounding DINO exhibits superior mAP50 and mAP compared to GLIP, OV-DETR, and YOLO-World. In this embodiment, the further improved GroundingDINO achieves improvements of 1.2% and 0.8% in mAP50 and mAP, respectively. In the more crucial experimental verification for unknown categories, Grounding DINO also demonstrates superior performance, with further improvements to mAP50 and mAP of 1.1% and 0.8%, respectively. This embodiment's improved Grounding DINO model shows advantages in both known and unknown categories, demonstrating the superiority of the improved Grounding DINO model.

[0044] The second-stage metal fitting corrosion detection algorithm was evaluated and trained based on a metal fitting corrosion target detection dataset. The improved YOLO model of this embodiment was compared with the metal fitting corrosion detection network composed of existing target detection models YOLOv5, YOLOv7, Gold-YOLO, YOLOv8, YOLOv9, YOLOv10, RT-DETR, and YOLOv11. The experimental results are shown in Table 3.

[0045] Table 3: Comparative Experimental Results of the Improved YOLO Model

[0046] The experimental results shown in Table 3 demonstrate that the method in this embodiment has significantly fewer parameters than other methods, reducing the number of parameters by 2M compared to the minimum YOLOv11, while also having lower computational cost (FLOPs). Compared to the baseline method YOLOv11, it improves mAP50 by 1.1% (92.4% vs 93.5%) and mAP by 0.9% (73.9% vs 74.8%). Compared to the current advanced YOLOv8, YOLOv9, YOLOv10, and YOLOv11 models, the improved YOLO model in this embodiment significantly enhances the detection accuracy of metal corrosion on the surface of power fittings.

[0047] Table 4: Comparative Experimental Results of the Improved YOLO Model

[0048] As shown in Table 4, the experimental results reveal that traditional object detection algorithms struggle to significantly improve detection performance, even with sufficiently high image resolutions (1024×1024), resulting in poor performance in detecting corrosion on metal surfaces. In contrast, the two-stage metal corrosion detection method, using a 640×640 input, achieves a substantial performance improvement over single-stage algorithms. Compared to the single-stage method YOLOv11, it achieves a 22.1% improvement in mAP50 (87.8% vs 65.7%) and a 20.0% improvement in mAP (72.2% vs 52.2%).

[0049] To verify the contribution of each module in the method of this embodiment to the improvement of the overall model detection accuracy, the first stage of the improved Grounding-DINO model gradually adds a dual-branch multi-scale structure sensing module (DMSM) and scale-adaptive loss (SAFit Loss) to the Grounding-DINO model. Detailed experimental results are shown in Tables 5 and 6.

[0050] Table 5: Known ablation-like experimental results of the improved Grounding-DINO model

[0051] Table 6: Experimental results of unknown class ablation in the improved Grounding-DINO model

[0052] Referring to Tables 5 and 6, after the dual-branch multi-scale structure perception module (DMSM), although dilated convolution expands the receptive field, relying solely on dilated convolution reduces local feature extraction. Therefore, the dual-branch approach, consisting of standard convolution and dilated convolutions with different expansion rates, can compensate for each other and has a significant advantage. Furthermore, adding spatial attention and channel attention after the merged features increases the relevance of channel features and improves positional awareness. This improves the model's mAP50 by 0.6% and mAP by 0.4% for known classes, and by 0.4% and 0.2% for unknown classes. Subsequently, adding scale-adaptive loss not only combines the advantages of NWD loss but also provides more stable and smooth training for objects of different sizes, thus making the detector exhibit higher robustness. Simultaneously, the model improves the mAP50 by 0.6% and mAP by 0.4% for known classes, and by 0.7% and 0.6% for unknown classes.

[0053] In the second stage of this embodiment, windmill-shaped convolution (PConv) and scale-aware contextual pyramid pooling (SCPP) modules were gradually added to the YOLOv11 model. To verify the effectiveness of the added modules, ablation experiments were conducted. These two modules were gradually added to the baseline YOLOv11 model to verify the detection effect of metal fitting corrosion. The results are shown in Tables 7 and 8.

[0054] Table 7: Ablation Experiment Results of Windmill-Shaped Convolution PConvc

[0055] Table 8: Ablation Experiment Results of the Improved YOLOv11 Model

[0056] The experimental results in Tables 7 and 8 show that increasing the PConv kernel size in the power fitting corrosion dataset does not bring additional performance gains. The windmill-shaped convolution PConv kernel size of 3 provides a more effective receptive field, which is crucial for capturing features of small targets. Since the feature map and target size decrease during downsampling, a kernel size of 3 already provides sufficient performance. Therefore, using the windmill-shaped convolution PConv with a kernel size of 3 is the most balanced choice, maintaining performance while reducing computational cost. Meanwhile, in the ablation experiments in Table 8, adding the windmill-shaped convolution PConv module improved the model's mAP50 by 0.5% and mAP by 0.4%. Compared with the C2PSA module, the introduction of the scale-aware contextual pyramid pooling module SCPP can better adaptively fuse features, resulting in a larger receptive field and significantly improved hardware corrosion detection performance. The experiments show that the scale-aware contextual pyramid pooling module SCPP alleviates the hole phenomenon and also reduces false positives and false negatives, improving the model's mAP50 by 0.6% and mAP by 0.5%.

[0057] In summary, this embodiment presents a power fitting corrosion detection method that integrates open vocabulary and structure awareness. Firstly, by improving the Grounding-DINO model, it enables the detection of various known power fittings and even more unknown ones. Leveraging the characteristics of open vocabulary detection, it covers as many possible fitting types as possible. By adding a dual-branch multi-scale structure awareness module, it overcomes the constraint of the fixed receptive field of traditional convolutional operations, improving its structure awareness capability and effectively establishing semantic associations for structured power fittings. Furthermore, scale-adaptive loss is used to enhance the network's adaptability to fitting scales. Secondly, in the second stage, the YOLOv11 model is improved by adding a windmill-shaped convolution PConv to the network backbone, enhancing the network's feature extraction capability for metal corrosion at different scales. The scale-aware context pyramid pooling module SCPP further alleviates the problem of high intra-class variability in fitting corrosion detection, which makes object features difficult to distinguish and easily leads to false detections and false negatives.

[0058] Furthermore, this embodiment also provides a power fitting corrosion detection system that integrates open vocabulary and structure awareness, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the power fitting corrosion detection method that integrates open vocabulary and structure awareness.

[0059] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0060] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting corrosion of power fittings that integrates open vocabulary and structure awareness, characterized in that, The process includes the following steps: S1, fine-tuning the backbone network using an image-text dataset from the power industry, wherein the backbone network includes an image encoder and a text encoder; S2, embedding the fine-tuned backbone network into a power fitting detection network and a fitting corrosion detection network, wherein the power fitting detection network is used to locate the position of power fittings from the input image based on the input text features, and the fitting corrosion detection network is used to detect corrosion from the power fitting image. The fitting detection network is trained using an open-vocabulary fine-tuning dataset, and the corrosion detection network is trained using a closed-set corrosion dataset. The open-vocabulary fine-tuning dataset includes bounding boxes of objects and corresponding text descriptions, while the closed-set corrosion dataset contains only corrosion images and corrosion category information; S3, using the trained fitting detection network to detect power fitting images from the input image and its corresponding text features, wherein the text features are tensors obtained by feature extraction of text through a trained text encoder, and the text features have a one-to-one correspondence with the object category in the input image. Based on the relationship, the detected power fitting images are used to complete corrosion detection using a trained fitting corrosion detection network. The power fitting detection network includes a dual-branch multi-scale structure perception module, a feature enhancement module, a language-guided query selection module, and a cross-modal decoder. The dual-branch multi-scale structure perception module is used to perform structure perception on the image features extracted by the image encoder. The feature enhancement module is used to enhance the features of the structure-perceived image features and the text features extracted by the text encoder. The language-guided query selection module is used to perform language-guided queries on the feature-enhanced image features and text features to obtain query features. The cross-modal decoder is used to perform cross-modal decoding based on the feature-enhanced image features, text features, and query features to obtain the localization box and confidence score of the power fitting. The dual-branch multi-scale structure perception module uses a dual-branch structure consisting of two branches to extract multi-scale features, and then passes the multi-scale features sequentially through a spatial attention module and a channel attention module to obtain the structure-perceived image features.

2. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The backbone network mentioned in step S1 is the Grounding DINO backbone network. The fine-tuning includes using an image-text dataset and a contrast loss function from the power field, and using LoRA technology to fine-tune the Grounding DINO backbone network.

3. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The dual-branch multi-scale structure perception module has two branches. The first branch includes two cascaded 3×3 combined convolutional modules (CBR3), each consisting of a 3×3 convolutional module, a batch normalization layer (BN), and a ReLU activation unit. The second branch includes two cascaded combined dilated convolutional modules (DCBR), each consisting of a dilated convolution, a batch normalization layer (BN), and a ReLU activation unit. The dilation rates of the dilated convolutional modules in the two DCBR modules are different. The functional expression for extracting multi-scale features using the dual-branch structure is as follows: ； in, For multi-scale features, It is a 1×1 combined convolutional module, which consists of a 3×3 convolutional module CBR, a batch normalization layer BN, and a ReLU activation unit; It is a 3×3 combined convolution module. For input features, For the combined dilated convolutional module DCBR with a dilation rate of 2, It is a combined dilated convolutional module DCBR with a dilation rate of 4.

4. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The feature enhancement module consists of multiple stacked enhancement layers with identical structures. The enhancement layers perform feature enhancement on the image features after structure perception and the text features extracted by the text encoder, including: first, performing intra-modal enhancement on the image features and text features respectively through parallel self-attention; then, performing inter-modal fusion on the intra-modal enhanced image features and text features through cross-attention to exchange semantic information between modalities; and finally, outputting the inter-modal fused image features and text features after nonlinear transformation by a feedforward network.

5. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The language-guided query selection module is used to perform language-guided queries on the enhanced image features and text features to obtain query features. This includes: calculating the dot product similarity between the image features and text features respectively to obtain a similarity matrix composed of dot product similarity; for each image feature position in the image features, taking the maximum value of the similarity with all text features in the similarity matrix to form a score vector representing the overall relevance between the image feature position and the text features; finally, selecting the top k image feature positions with the highest scores from the score vector and extracting the corresponding image features based on these position indices as the obtained query features.

6. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The cross-modal decoder includes multiple stacked decoding layers and a detection head. The decoding layers perform cross-modal decoding based on enhanced image features, text features, and query features to obtain images of power fittings. This includes: extracting target information of interest for each of the input image features, text features, and query features through self-attention layers to obtain enhanced image features, text features, and query features; enabling intermodal interaction between query features and text features through cross-attention layers to fuse language semantics for target localization; enabling intermodal interaction between query features and image features after intermodal interaction through cross-attention layers to capture visual details; outputting the query features after completing two intermodal interactions through a feedforward network for feature transformation, and then using the detection head to regress and classify the features output by the last decoding layer to generate bounding boxes and category prediction results.

7. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, In step S2, when training the hardware detection network using an open-vocabulary fine-tuning dataset and the corrosion detection network using a closed-set corrosion dataset, the loss function used includes scale-adaptive loss. The expression for the calculation function of the scale-adaptive loss is as follows: ； ； ； ； in, For scale-adaptive loss, For scale adaptation, It is a natural constant. To predict the area of ​​the bounding box, For constant hyperparameters, The intersection-union ratio (IUU) of the predicted bounding box and the target bounding box. For constant hyperparameters Normalized Wasserstein distance, For constant hyperparameters Normalized Wasserstein distance, For prediction boxes and target box The Wasserstein distance between Gaussian distributions For prediction boxes and target box The L2 norm.

8. The method for detecting corrosion of power fittings integrating open vocabulary and structure awareness according to claim 1, characterized in that, The metal fitting corrosion detection network is an improved YOLOv11 model, which consists of a target detection backbone network, a neck network, and a head network. The target detection backbone network consists of two cascaded convolutional layers, four feature extraction units, a fast spatial pyramid pooling module (SPPF), and a scale-aware context pyramid pooling module (SCPP). The feature extraction units consist of an improved C3K2 module and a convolutional layer. The improved C3K2 module is obtained by replacing the 3×3 convolutional module in the bottleneck layer (Bottleneck) of the original C3K2 module with a windmill-shaped convolutional module (PConv). The scale-aware context pyramid pooling module (SCPP) includes three parallel branches, where the first branch is used to process the input feature F. input The first output feature is obtained through a 1×1 convolution, and the second branch converts the input feature F... input The second output feature is obtained by using a multi-scale module, and the third branch extracts the input feature F. input Global average pooling is used to obtain weights. The first and second output features are concatenated by the channel dimension and then multiplied by the weights. This allows the output features of the Scale-Aware Context Pyramid Pooling (SCPP) module to be obtained through a 1×1 convolution. The multi-scale module includes three parallel sub-branches, each of which includes a dilated convolution, a compression activation module, and a 1×1 convolution. The dilation rates of the dilated convolutions in the three sub-branches are different. The feature maps output by the 1×1 convolutions of the three sub-branches are summed and then subjected to a 3×3 convolution and a Softmax function to generate pixel-level attention maps of three scales. , and The second output feature is calculated according to the following formula. : ； in, For element-wise multiplication, , and These are the feature maps output by the 1×1 convolution of the three sub-branches.

9. A corrosion detection system for power fittings integrating open vocabulary and structure awareness, comprising an interconnected microprocessor and memory, characterized in that, The microprocessor is programmed or configured to execute the power fitting corrosion detection method that integrates open vocabulary and structure awareness as described in any one of claims 1 to 8.

Images