Large device surface corrosion detection method and system

Abstract

The invention discloses a method and a system for detecting surface corrosion of a large-scale device, which are used for solving the technical problem of poor target detection accuracy caused by the existing surface corrosion detection technology of the large-scale device. The method comprises the steps that a corrosion image of a to-be-detected large-scale device is acquired and input into a preset large-scale device surface corrosion detection model, and the preset large-scale device surface corrosion detection model comprises a fusion attention feature extraction network, a target neck network and a target detection head; performing feature enhancement extraction on the corrosion image of the to-be-detected large-scale device by adopting the fusion attention feature extraction network to generate an enhanced feature map; performing feature convolution fusion on the enhanced feature map through the target neck network, and outputting a fused feature map; and inputting the fused feature map into a target detection head for target detection, and outputting a corrosion detection result.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a large-scale device surface corrosion detection method and system. Background Art

[0002] Corrosion morphology image features are the basis for corrosion morphology recognition and classification, including grayscale (density, color) features, texture features and corrosion shape features, such as the distribution of corrosion points, the size of corrosion points and the corrosion area. The recognition of corrosion morphology features is the difficulty of corrosion image processing and quantitative evaluation.

[0003] With the advancement of machine learning and machine vision technology, a lot of research has been done on vision-based structural corrosion and coating defect detection and identification. Traditional machine learning algorithms based on classifiers such as decision trees, random forests, and support vector machines (SVM) have made great progress. Huynh et al. proposed a multi-class support vector machine (SVM) model for evaluating coating conditions, and its efficient corrosion damage identification performance has been verified by high-definition images collected from the Sydney Harbour Bridge.

[0004] Most of the existing large-scale equipment surface corrosion detection technologies use continuous wavelet transform to decompose images, and then extract energy values ​​as feature information to further analyze the correlation between image feature values ​​and corrosion weight loss data. However, in the face of images with severe corrosion or high noise, the detailed features of the corrosion image are easily disturbed and become blurred, which makes it impossible to effectively identify the key information in the image, resulting in poor accuracy of target detection. Summary of the invention

[0005] The present invention provides a large-scale device surface corrosion detection method and system, which are used to solve the technical problem that the existing large-scale device surface corrosion detection technology leads to poor accuracy of target detection.

[0006] A first aspect of the present invention provides a method for detecting surface corrosion of a large device, comprising:

[0007] Acquire a corrosion image of a large device to be detected, and input the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, wherein the preset large device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network, and a target detection head;

[0008] Using the fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected, and generate an enhanced feature map;

[0009] Performing feature convolution fusion on the enhanced feature map through the target neck network, and outputting a fused feature map;

[0010] The fused feature map is input into the target detection head for target detection, and the corrosion detection result is output.

[0011] Optionally, the fused attention feature extraction network includes a convolution fusion attention mechanism module and a multi-scale feature fusion module; the enhanced feature map includes a first enhanced feature map, a second enhanced feature map and a third enhanced feature map; the use of the fused attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected to generate an enhanced feature map includes:

[0012] Using a convolution fusion attention mechanism module to perform convolution fusion on the corrosion image of the large device to be detected to generate a first convolution fusion feature map;

[0013] Performing convolution fusion on the first convolution fusion feature map through a convolution fusion attention mechanism module, and outputting a first enhanced feature map;

[0014] Inputting the first enhanced feature map into a convolutional fusion attention mechanism module for convolutional fusion to determine a second enhanced feature map;

[0015] Using a convolution fusion attention mechanism module to perform convolution fusion on the second enhanced feature map, and output a second convolution fusion feature map;

[0016] The second convolution fusion feature map is subjected to multi-scale feature fusion through a multi-scale feature fusion module to generate a third enhanced feature map.

[0017] Optionally, the convolution fusion attention mechanism module includes two serially connected convolution layers, a convolution block attention mechanism layer, a splicing layer, and a convolution layer; the image processing step of the convolution fusion attention mechanism module includes:

[0018] The eroded input map of the input convolution fusion attention mechanism module is convolved through two series-connected convolution layers to generate the first convolution feature map;

[0019] Segmenting the first convolution feature map, and outputting a first segmentation feature map and a second segmentation feature map;

[0020] Using two convolutional layers connected in series to perform a convolution operation on the second segmentation feature map to generate a second convolutional feature map;

[0021] Adjusting the feature weight of the second convolution feature map through the convolution block attention mechanism layer, and outputting the first convolution block attention feature map;

[0022] Splicing the first convolution block attention feature map and the second segmentation feature map to generate a first spliced ​​feature map, and inputting the first spliced ​​feature map into two serially connected convolutional layers for convolution operation, and outputting a third convolution feature map;

[0023] Using a convolutional block attention mechanism layer to adjust the feature weight of the third convolutional feature map to generate a second convolutional block attention feature map;

[0024] Splicing the second convolution block attention feature map and the first splicing feature map, and outputting a second splicing feature map;

[0025] Splicing the first segmentation feature map, the first convolution block attention feature map, the second splicing feature map, and the second segmentation feature map to generate a third splicing feature map;

[0026] A convolution operation is performed on the third concatenated feature map through a convolution layer to output a corrosion output map.

[0027] Optionally, the multi-scale feature fusion module includes three cascaded average pooling layers, three cascaded maximum pooling layers, a splicing layer and a convolution layer; the multi-scale feature fusion module performs multi-scale feature fusion on the second convolution fusion feature map to generate a third enhanced feature map, including:

[0028] Performing a convolution operation on the second convolution fusion feature map through a convolution layer to generate an enhanced convolution feature map;

[0029] Inputting the enhanced convolutional feature map into three cascaded average pooling layers and three cascaded maximum pooling layers for average pooling and maximum pooling, respectively, and outputting an average pooling feature map and a maximum pooling feature map;

[0030] Using a splicing layer to splice the maximum pooling feature map, the average pooling feature map and the enhanced convolution feature map, and output a spliced ​​pooling feature map;

[0031] A convolution operation is performed on the concatenated pooled feature map through a convolution layer to generate a third enhanced feature map.

[0032] Optionally, the target neck network includes a partial convolution fusion module, a splicing layer, an upsampling layer, and a convolution layer; the fused feature map includes a first fused feature map, a second fused feature map, and a third fused feature map; and the step of performing feature convolution fusion on the enhanced feature map through the target neck network and outputting the fused feature map includes:

[0033] Upsampling the third enhanced feature map by using an upsampling layer to output a first upsampling feature map, and splicing the first upsampling feature map and the second enhanced feature map by using a splicing layer to generate a first intermediate splicing feature map;

[0034] Inputting the first intermediate splicing feature map into the partial convolution fusion module for partial convolution fusion, outputting a first partial convolution fusion feature map, and upsampling the first partial convolution fusion feature map using an upsampling layer to generate a second upsampled feature map;

[0035] Splicing the second up-sampled feature map and the first enhanced feature map through a splicing layer, outputting a second intermediate splicing feature map, inputting the second intermediate splicing feature map into a partial convolution fusion module for partial convolution fusion, and outputting a first fused feature map;

[0036] Using a convolution layer to perform a convolution operation on the first fused feature map to generate a first fused convolution feature map, and inputting the first fused convolution feature map and the first partial convolution fused feature map into a splicing layer for splicing, and outputting a third intermediate splicing feature map;

[0037] Performing partial convolution fusion on the third intermediate spliced ​​feature map through a partial convolution fusion module to generate a second fused feature map, and performing a convolution operation on the second fused feature map using a convolution layer to output a second fused convolution feature map;

[0038] The second fused convolution feature map and the third enhanced feature map are input into the splicing layer for splicing, a fourth intermediate splicing feature map is output, and the fourth intermediate splicing feature map is partially convolutionally fused through a partial convolution fusion module to generate a third fused feature map.

[0039] Optionally, the training process of the preset large-scale device surface corrosion detection model includes:

[0040] Acquire a corrosion image dataset of a large-scale device to be trained, and annotate the corrosion image dataset of the large-scale device to be trained to generate an annotated corrosion image dataset;

[0041] Using a preset mosaic method to combine the labeled corrosion image data set to generate a plurality of combined corrosion images;

[0042] Performing cropping and adjustment on each of the combined eroded images respectively to generate a plurality of adjusted eroded images;

[0043] Using a plurality of the adjusted corrosion images to construct a target corrosion image data set;

[0044] The target corrosion image data set is used to perform model training on an initial large-scale device surface corrosion detection model, and the trained preset large-scale device surface corrosion detection model is determined.

[0045] A second aspect of the present invention provides a large-scale device surface corrosion detection system, comprising:

[0046] An acquisition module, used for acquiring a corrosion image of a large device to be detected, and inputting the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, wherein the preset large device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network and a target detection head;

[0047] A feature extraction module, used to use the fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected, and generate an enhanced feature map;

[0048] A convolution fusion module, used for performing feature convolution fusion on the enhanced feature map through the target neck network, and outputting a fused feature map;

[0049] The target detection module is used to input the fused feature map into the target detection head for target detection and output the corrosion detection result.

[0050] A third aspect of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large-scale device surface corrosion detection method as described in any one of the above items.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the large-scale device surface corrosion detection method as described in any one of the above items.

[0052] A fifth aspect of the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer executes the steps of the large-scale device surface corrosion detection method as described in any one of the above items.

[0053] It can be seen from the above technical solutions that the present invention has the following advantages:

[0054] The above technical scheme of the present invention provides a method for detecting surface corrosion of large-scale devices. First, a corrosion image of a large-scale device to be detected is obtained, and the corrosion image of the large-scale device to be detected is input into a preset surface corrosion detection model of a large-scale device, wherein the preset surface corrosion detection model of a large-scale device includes a fusion attention feature extraction network, a target neck network and a target detection head; then, a fusion attention feature extraction network is used to perform feature enhancement extraction on the corrosion image of the large-scale device to be detected, and an enhanced feature map is generated; feature convolution fusion is performed on the enhanced feature map through the target neck network, and a fusion feature map is output; finally, the fusion feature map is input into the target detection head for target detection, and a corrosion detection result is output; based on the above scheme, after feature enhancement extraction and feature convolution fusion are performed on the acquired corrosion image of the large-scale device to be detected through the fusion attention feature extraction network and the target neck network, the obtained fusion feature map is input into the target detection head for target detection, and the process of outputting the corrosion detection result can automatically focus on the significant areas in the image, enhance the ability to capture key information, and thus improve the accuracy of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 A flowchart of the steps of a large-scale device surface corrosion detection method provided in Example 1 of the present invention;

[0057] Figure 2 A schematic diagram of the structure of a pre-set large-scale device surface corrosion detection model provided in the first embodiment of the present invention;

[0058] Figure 3 A schematic diagram of the structure of a C2f-CBAM module provided in Embodiment 1 of the present invention;

[0059] Figure 4 A schematic diagram of the structure of a multi-scale feature fusion module provided in Embodiment 1 of the present invention;

[0060] Figure 5 A flowchart of the steps of model training for presetting a large-scale device surface corrosion detection model provided in the second embodiment of the present invention;

[0061] Figure 6 A schematic diagram of a flow chart of model training for a pre-set large-scale device surface corrosion detection model provided in the second embodiment of the present invention;

[0062] Figure 7A structural block diagram of a surface corrosion detection system for a large device is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The embodiments of the present invention provide a large-scale device surface corrosion detection method and system, which are used to solve the technical problem that the existing large-scale device surface corrosion detection technology leads to poor accuracy of target detection.

[0064] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] See also Figure 1 , Figure 1 A flowchart of the steps of a large-scale device surface corrosion detection method provided in Example 1 of the present invention.

[0066] The present invention provides a large-scale device surface corrosion detection method, comprising:

[0067] Step 101, obtain the corrosion image of the large device to be detected, and input the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, which includes a fusion attention feature extraction network, a target neck network and a target detection head.

[0068] It should be noted that the present invention introduces the CBAM (Convolutional Block Attention Module) attention mechanism to the C2f (Cross Stage Partial Bottleneck with 2 Convolutions) model in the YOLOv8 model, increases the channel and spatial attention bias to the corrosion morphology feature information, and adds the corresponding weight matrix to the target feature map by fusing the attention mechanism to achieve fine features; based on the traditional SPPF (Spatial Pyramid Pooling Fast) module of the YOLOv8 model, an average pooling layer is added so that the feature information can be fused and integrated at different spatial scales. Compared with the feature extraction of a single scale, multi-scale feature fusion can better handle targets of different sizes, shapes and positions in the image, thereby constructing a target detection model based on the improved YOLOv8, that is, a preset large-scale device surface corrosion detection model.

[0069] See also Figure 2The preset large-scale device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network and a target detection head; the fusion attention feature extraction network is composed of four convolutional fusion attention mechanism modules and a multi-scale feature fusion module, and the convolutional fusion attention mechanism module is composed of a convolutional layer and a C2f-CBAM module in series; the target neck network is composed of four partial convolutional fusion modules (C2f), four concatenation layers (Concat), two upsampling layers (Upsample), and two convolutional layers.

[0070] Step 102: Use a fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be inspected, and generate an enhanced feature map.

[0071] The fusion attention feature extraction network includes a convolutional fusion attention mechanism module and a multi-scale feature fusion module. Figure 3 The convolutional fusion attention mechanism module consists of a convolutional layer and a C2f-CBAM module in series. The C2f-CBAM module consists of six convolutional layers, two convolutional block attention mechanism layers, and three splicing layers.

[0072] For further information, see Figure 4 ,The multi-scale feature fusion module consists of two convolutional layers, three cascaded average pooling layers, three cascaded maximum pooling layers and a concatenation layer.

[0073] The enhanced characteristic map includes a first enhanced characteristic map, a second enhanced characteristic map and a third enhanced characteristic map.

[0074] Further, step 102 may include the following sub-steps S21-S25:

[0075] Step S21, using a convolution fusion attention mechanism module to perform convolution fusion on the corrosion image of the large device to be detected, to generate a first convolution fusion feature map;

[0076] Optionally, the convolution fusion attention mechanism module includes two serially connected convolution layers, a convolution block attention mechanism layer, a splicing layer, and a convolution layer; the image processing steps of the convolution fusion attention mechanism module include:

[0077] S01, convolve the eroded input map of the input convolution fusion attention mechanism module through two series-connected convolution layers to generate a first convolution feature map;

[0078] S02, segmenting the first convolution feature map, and outputting a first segmentation feature map and a second segmentation feature map;

[0079] S03, using two serially connected convolutional layers to perform a convolution operation on the second segmentation feature map to generate a second convolutional feature map;

[0080] S04, adjusting the feature weight of the second convolution feature map through the convolution block attention mechanism layer, and outputting the first convolution block attention feature map;

[0081] S05, concatenating the attention feature map of the first convolution block and the second segmentation feature map to generate a first concatenated feature map, and inputting the first concatenated feature map into two serially connected convolutional layers for convolution operation, and outputting a third convolutional feature map;

[0082] S06. Use the convolution block attention mechanism layer to adjust the feature weight of the third convolution feature map to generate a second convolution block attention feature map;

[0083] S07, concatenating the attention feature map of the second convolution block and the first concatenated feature map, and outputting a second concatenated feature map;

[0084] S08, splicing the first segmentation feature map, the first convolution block attention feature map, the second splicing feature map, and the second segmentation feature map to generate a third splicing feature map;

[0085] S09. Perform a convolution operation on the third concatenated feature map through a convolution layer, and output a corrosion output map.

[0086] The corrosion input map is a feature map of the input convolution fusion attention mechanism module. It can be understood that the corrosion input map can correspond to any feature map of the input convolution fusion attention mechanism module for image processing during model training or corrosion detection;

[0087] The first convolution feature map, the first segmentation feature map, the second segmentation feature map, the second convolution feature map, the first convolution block attention feature map, the first splicing feature map, the third convolution feature map, the second convolution block attention feature map, the second splicing feature map, and the third splicing feature map are all intermediate maps generated in the convolution fusion attention mechanism module;

[0088] The eroded output image is a feature map output by the convolutional fusion attention mechanism module. It can be understood that it can correspond to any feature map output after the convolutional fusion attention mechanism module performs image processing during model training or detection.

[0089] It should be noted that the C2f module is a component in the YOLOv8 model for feature fusion. Its role is to fuse feature maps at different stages to obtain higher quality feature representations. This cross-stage fusion helps improve the performance of target detection. CBAM (Convolutional Block Attention Mechanism Layer) is an attention mechanism used to enhance the performance of convolutional neural networks. The convolutional block attention mechanism layer consists of two parts: channel attention and spatial attention. The channel attention module is used to adaptively adjust the weights of feature maps of different channels to highlight important features. The spatial attention module is used to adaptively adjust the weights of feature maps at different spatial locations to improve the clarity of feature maps.

[0090] Based on the above foundation, the present invention combines C2f and CBAM attention mechanisms. C2f fused with CBAM attention mechanism takes C2f fused feature map as input, and further enhances the representation ability of feature map through CBAM attention mechanism. Therefore, the C2f module of Backbone network in YOLOv8 is replaced with C2f-CBAM module. This combination helps to improve the performance of target detection model, so that it can more effectively capture important features in corrosion images and improve the accuracy of target detection.

[0091] Step S22: performing convolution fusion on the first convolution fusion feature map through a convolution fusion attention mechanism module, and outputting a first enhanced feature map;

[0092] Step S23: input the first enhanced feature map into the convolution fusion attention mechanism module for convolution fusion to determine the second enhanced feature map;

[0093] Step S24: using a convolution fusion attention mechanism module to perform convolution fusion on the second enhanced feature map, and output a second convolution fusion feature map;

[0094] Step S25: Perform multi-scale feature fusion on the second convolution fusion feature map through a multi-scale feature fusion module to generate a third enhanced feature map.

[0095] The multi-scale feature fusion module consists of three cascaded average pooling layers, three cascaded maximum pooling layers, a concatenation layer, and a convolutional layer.

[0096] Further, step S25 may include the following sub-steps S251-S254:

[0097] Step S251, performing a convolution operation on the second convolution fusion feature map through a convolution layer to generate an enhanced convolution feature map;

[0098] Step S252: input the enhanced convolution feature map to three cascaded average pooling layers and three cascaded maximum pooling layers for average pooling and maximum pooling, respectively, and output the average pooling feature map and the maximum pooling feature map;

[0099] Step S253: Use a splicing layer to splice the maximum pooling feature map, the average pooling feature map and the enhanced convolution feature map, and output a spliced ​​pooling feature map;

[0100] Step S254: performing a convolution operation on the concatenated pooled feature map through a convolution layer to generate a third enhanced feature map.

[0101] It should be noted that SPPF is different from the traditional SPP, which uses multiple pooling kernels of different sizes. SPPF usually uses multiple small pooling layers (such as three layers of 3×3 maximum pooling) to replace the pooling operation of a single large kernel. This strategy reduces the amount of calculation while maintaining the ability to fuse multi-scale features. After the fast pooling operation, the SPPF layer concatenates the outputs of different pooling layers to form a feature vector of fixed length. This feature vector can then be used as the input of subsequent fully connected layers or other network layers.

[0102] Furthermore, in order to further enhance the feature extraction capability of the SPPF module, the present invention proposes a multi-scale feature fusion scheme. The core idea of ​​this scheme is to enhance the network's perception of multi-scale targets by capturing the feature information of the image at different scales. Specifically, on the basis of the traditional SPPF, an average pooling layer is added so that the feature information can be fused and integrated at different spatial scales. Compared with single-scale feature extraction, multi-scale feature fusion can better handle targets of different sizes, shapes and positions in the image, and improve the network's expression ability and adaptability to complex scenes.

[0103] Furthermore, max pooling extracts the maximum value from the feature map of the local receptive field. It retains the most significant features, usually strong responses to edges, textures, or significant target areas. Average pooling generates a smoother feature map by calculating the average of all values ​​in the local receptive field. It retains the overall information in the local area instead of just focusing on the maximum value. By introducing the average pooling layer, not only can the key information be effectively retained, but also the information loss caused by the pooling operation can be reduced, thereby achieving a more efficient and accurate feature representation. The combination of the average pooling layer and the maximum pooling layer can integrate different statistical information, further improve the model's ability to capture detailed features, and provide richer and more recognizable feature maps for downstream target detection tasks. At the same time, this solution improves the generalization performance of the model while maintaining computational efficiency, and is suitable for various application scenarios with limited computing resources.

[0104] Step 103: Perform feature convolution fusion on the enhanced feature map through the target neck network, and output the fused feature map.

[0105] The target neck network includes a partial convolution fusion module, a splicing layer, an upsampling layer, and a convolution layer; the fused feature map includes a first fused feature map, a second fused feature map, and a third fused feature map.

[0106] The fused feature map includes a first fused feature map, a second fused feature map and a third fused feature map.

[0107] Further, step 103 may include the following sub-steps S31-S36:

[0108] Step S31, upsampling the third enhanced feature map by using an upsampling layer to output a first upsampling feature map, and splicing the first upsampling feature map and the second enhanced feature map by using a splicing layer to generate a first intermediate splicing feature map;

[0109] Step S32: input the first intermediate splicing feature map into the partial convolution fusion module for partial convolution fusion, output the first partial convolution fusion feature map, and use the upsampling layer to upsample the first partial convolution fusion feature map to generate a second upsampled feature map;

[0110] Step S33, splicing the second up-sampled feature map and the first enhanced feature map through the splicing layer, outputting a second intermediate splicing feature map, inputting the second intermediate splicing feature map into the partial convolution fusion module for partial convolution fusion, and outputting a first fused feature map;

[0111] Step S34: using a convolution layer to perform a convolution operation on the first fused feature map to generate a first fused convolution feature map, and inputting the first fused convolution feature map and the first partial convolution fused feature map into a splicing layer for splicing, and outputting a third intermediate splicing feature map;

[0112] Step S35, performing partial convolution fusion on the third intermediate splicing feature map through the partial convolution fusion module to generate a second fused feature map, and performing a convolution operation on the second fused feature map using a convolution layer to output a second fused convolution feature map;

[0113] Step S36: input the second fused convolution feature map and the third enhanced feature map into the splicing layer for splicing, output a fourth intermediate splicing feature map, and perform partial convolution fusion on the fourth intermediate splicing feature map through a partial convolution fusion module to generate a third fused feature map.

[0114] Step 104: input the fused feature map to the target detection head for target detection, and output the corrosion detection result.

[0115] The corrosion detection results include a first corrosion detection result, a second corrosion detection result, and a third corrosion detection result.

[0116] It should be noted that the first fused feature map, the second fused feature map and the third fused feature map are respectively input into the corresponding target detection head for target detection, and the first corrosion detection result corresponding to the first fused feature map, the second corrosion detection result corresponding to the second fused feature map and the third corrosion detection result corresponding to the third fused feature map are output.

[0117] In an embodiment of the present invention, the present invention provides a large-scale device surface corrosion detection method, first, a corrosion image of a large-scale device to be detected is obtained, and the corrosion image of the large-scale device to be detected is input into a preset large-scale device surface corrosion detection model, the preset large-scale device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network and a target detection head; then, the fusion attention feature extraction network is used to perform feature enhancement extraction on the corrosion image of the large-scale device to be detected, and an enhanced feature map is generated; the enhanced feature map is subjected to feature convolution fusion through the target neck network, and a fused feature map is output; finally, the fused feature map is input into the target detection head for target detection, and a corrosion detection result is output; based on the above scheme, after the feature enhancement extraction and feature convolution fusion are performed on the acquired corrosion image of the large-scale device to be detected through the fusion attention feature extraction network and the target neck network, the obtained fused feature map is input into the target detection head for target detection, and the process of outputting the corrosion detection result can automatically focus on the significant area in the image, enhance the ability to capture key information, and thus improve the accuracy of target detection.

[0118] For better explanation, refer to Figure 5 , shows a flowchart of the steps of model training of a preset large-scale device surface corrosion detection model provided in the second embodiment of the present invention, and the process may include the following steps:

[0119] Step 501: Obtain a corrosion image dataset of a large-scale device to be trained, and annotate the corrosion image dataset of the large-scale device to be trained to generate an annotated corrosion image dataset.

[0120] It should be noted that target annotation is performed on the acquired image, that is, the corroded portion of the steel in the image is annotated, so as to ensure that the newly generated data has the same label information as the original data.

[0121] Step 502: Use a preset mosaic method to combine the labeled corrosion image data set to generate a plurality of combined corrosion images.

[0122] It should be noted that after collecting a series of steel corrosion pictures as raw data and annotating them, four pictures are randomly selected from them, and based on the preset mosaic method, these four pictures are randomly combined into a large mosaic image (combined corrosion image). These four pictures can be placed in four different areas of the mosaic image, and ensure that they have a certain overlap so that the model can learn the continuity between different parts.

[0123] In this embodiment, the Mosaic method is used to expand the steel corrosion image dataset to increase the diversity and complexity of the data, thereby improving the generalization ability and performance of the model.

[0124] Step 503: perform cropping and adjustment on each combined erosion image to generate a plurality of adjusted erosion images.

[0125] It should be noted that random cropping and adjustment are performed on the generated Mosaic image. The cropping position and cropping size are randomly selected, and the cropped image is resized, and the brightness and contrast are adjusted to increase the diversity of the data.

[0126] Step 504: Use multiple adjusted erosion images to construct a target erosion image data set.

[0127] It should be noted that the cropped and adjusted Mosaic image (adjusted eroded image) and the corresponding annotation information in the image are saved as new training data, namely, the target eroded image data set.

[0128] Step 505: Use the target corrosion image data set to perform model training on the initial large-scale device surface corrosion detection model to determine the trained preset large-scale device surface corrosion detection model.

[0129] It should be noted that in the target detection task, the model presets a large number of prior frames for each image. During the training process, the network will update the weight parameters according to the overlap between the predicted area and the actual area, that is, the IOU of the two areas. In order to evaluate the prediction results and compare with other methods, the present invention uses IOU to quantitatively evaluate the performance of the detection model (the trained preset large-scale device surface corrosion detection model). IOU is defined as follows:

[0130] ;

[0131] Among them, G T is the ground truth, i.e. the marked target box; IOU (G T ,P M) is the evaluation index, which can be set to different sizes. When the IOU between the predicted area and the true area is greater than the set value of 0.5, the predicted frame is regarded as a true prediction, called a positive sample, otherwise it is a negative sample; P M represents the predicted bounding box (Predicted Box), which is the detection box output by the model.

[0132] Furthermore, in the evaluation phase, precision is defined as the ratio of the number of positive samples to the number of recognized samples. Recall is defined as the ratio of correctly recognized objects to the total number of such objects in the test dataset. Precision and recall are defined as follows:

[0133] ;

[0134] Among them, TP is positive sample detection (True Positive); FN is negative sample (False Negative), which is detected incorrectly; FP is positive sample, which is detected incorrectly; Recall is the recall rate; Precision is the precision rate.

[0135] Furthermore, accuracy and recall are often contradictory and insufficient to measure model performance alone. Therefore, the present invention introduces AP (average precision) to intuitively represent the performance of the predictor. It indicates the performance of the model on objects of a specific category. The value of AP is the area enclosed by the PR (Precision-Recall) curve and the coordinate axis. However, AP represents the accuracy of a certain category and does not measure the overall performance of the entire model. In order to comprehensively evaluate the model performance and compare it with mainstream models, the present invention introduces mAP (mean Average Precision) as an evaluation indicator. mAP represents the average accuracy of each category of the model in the data set, which can well reflect the comprehensive performance of the model. AP and mAP are defined as follows:

[0136] ;

[0137] Among them, AveP(q) is the average accuracy of the qth category; Q is the total number of categories.

[0138] For example, see Figure 6, based on the image acquisition device, the original image set is obtained, the Mosaic method is used on the original image set, 4 images are randomly selected to generate new images, until all images are stitched together, and a data set with higher data diversity and model robustness is constructed; the image enhanced data set is divided into a training set and a validation set in an 8:2 ratio; the part of the data set that is divided into the training set is used to train the improved corrosion detection model to obtain the improved model, and the validation data set is used to evaluate the trained model. A comparative experiment is conducted on the baseline YOLOv8 model and the improved model to obtain the final steel corrosion detection model.

[0139] As a comparison of technical effects, it can be combined with existing technologies for reference. In corrosion science, corrosion digital image processing technology was first used to identify and evaluate stainless steel pitting. In 1981, D. Zhak et al. used image methods to study the corrosion image of AISI304 stainless steel in 10% FeC13. The image was binarized and the pitting rate value of 9.73% was obtained.

[0140] Kurose et al. used image analysis technology to analyze the distribution and development process of stress corrosion cracks in pure copper, proving that fractal dimension is an effective method for evaluating the stage of copper corrosion. Fractals began to be used in the extraction of corrosion morphology image features. Wang Shouyan et al. used the fractal dimension, generalized dimension, area factor and void features of the corrosion specimen image as input, and established a corrosion morphology image diagnosis system based on analytical features using fuzzy pattern recognition theory, and judged the surface corrosion morphology of the metal by the fractal feature value of the image.

[0141] Song Shizhe, Wang Shouyan and others used continuous wavelet transform to decompose images and extract energy values ​​as feature information, analyzed the correlation between image feature values ​​and corrosion weight loss data, and proved that this method can effectively judge and predict the early atmospheric corrosion behavior of pure zinc. Mohammad et al. used wavelet analysis technology to perform two-dimensional wavelet transform on images with pits and cracks, decomposed them into sub-images with various special scales and orientation information, extracted energy features, and applied neural network methods to achieve better corrosion morphology classification. Tao Lei et al. used wavelet image analysis technology to decompose the filtered LC4CS and LY12CZ aluminum alloy corrosion morphology images and extract the image feature values ​​δ of the sub-images to analyze the influence of various test factors on the test acceleration. Yan Yunhui et al. used the L1 norm, energy, and entropy output of each frequency band after wavelet transform as metal fracture identification and classification characteristics, and used the linear minimum distance classifier to accurately identify and classify six typical fractures such as corrosion fatigue and ductile fatigue.

[0142] In recent years, with the rapid development of computer technology and the advent of the era of big data, deep learning technology has begun to develop rapidly. The convolutional neural network trained by the deep learning classification algorithm has shown high efficiency and accuracy in image recognition and feature extraction. The reason is that the deep convolutional neural network has added convolutional layers and pooling layers compared to traditional neural networks, and can train and learn the features we need from high-dimensional output features, which not only reduces the computational workload but also greatly improves the classification accuracy, making it have a unique advantage in image classification processing. The convolutional neural network models such as LeNet, AleXNet, GoogLeNet, and VGGNet, which have won the championship and runner-up in the ILSVRC competition, have high accuracy in image classification and image recognition. Wang Minghao used deep learning methods to identify and retrieve the feature information of the microscopic morphology image of NiCrAlY coating / Ni-based high-temperature alloy. The RMS Prop optimizer was selected, and the recognition accuracy of the test set of the neural network with multiple convolutional layers was 98% and 90.67% respectively. The recognition accuracy of the convolutional neural network trained with the Adam optimizer was further improved to 99.17%. Wang built a pixel-based corrosion classifier by integrating a texture-based unsupervised image segmentation method with a classifier based on red, green and blue (RGB) features. Users only need to use a small set of manually labeled images to build a labeling tool that automatically labels corrosion-like features in images.

[0143] Therefore, deep learning, as a method that has been popular in recent years and has performed well, is used as the core method for studying corrosion image detection and recognition. A corrosion detection algorithm model with good performance and application prospects is proposed. Based on the current research difficulties in the field of corrosion image data, which are small and lacking in types, the impact of insufficient data and difficulty in data labeling in deep learning on the performance of the method model is overcome.

[0144] YOLOv8 is a real-time object detection model that is optimized for accuracy and speed, uses a more efficient network architecture, and can better handle objects in complex scenes. YOLOv8 can accurately identify and locate various types of corrosion features, such as rust, spalling, and cracks, which is very important for maintenance and repair. In many industrial scenarios, corrosion may appear in complex backgrounds. YOLOv8 performs well in handling this situation and can effectively distinguish corrosion features from background interference. However, high-quality and large amounts of annotated data are necessary for model training to ensure the accuracy of the model, especially in specific areas such as corrosion detection. And when dealing with very small corrosion features, the model may still be at risk of missing detection.

[0145] Based on the above foundations, existing technologies face many challenges in the field of corrosion image processing, the first of which is the serious shortage of corrosion image datasets. Most public datasets are small in scale, cover limited corrosion types and scenarios, and are difficult to fully reflect the complexity of actual industrial environments. Therefore, the existing datasets are not representative enough to cover the diversity of different materials, corrosion types, and environmental conditions, resulting in insufficient generalization of the model.

[0146] In addition, the lack of recognized standardized datasets in the field makes it difficult for researchers to unify the evaluation criteria for model performance. This lack of a unified benchmark makes it difficult to compare different methods horizontally and accurately measure the pros and cons of each method. At the same time, due to insufficient data, it is often difficult to fully train deep learning models, resulting in low accuracy in practical applications. Deep learning models usually rely on large amounts of data to extract features and patterns, but the existing scarce data cannot effectively support the model's learning process, limiting its ability to detect corrosion.

[0147] In summary, corrosion image processing technology urgently needs to solve the problems of small data set size, insufficient representativeness, and lack of standard benchmarks. Only by significantly improving the amount and quality of data can we promote the further application of deep learning models in this field, thereby improving the accuracy and reliability of corrosion detection.

[0148] In view of the above problems, the present invention proposes a method for detecting surface corrosion of large-scale devices, which makes full use of limited data sets and adopts means of feature fusion and enhanced attention mechanism to provide accurate and efficient detection of corrosion images. Specifically, by integrating attention mechanism and multi-scale feature fusion to achieve accuracy and feature extraction capability, the model can automatically focus on the significant areas in the image, thereby improving the ability to capture key information. At the same time, compared with the existing feature extraction methods, feature extraction methods often only process features of a single scale, which limits the model's effective recognition of targets of different sizes, shapes and distributions. The present invention adopts a multi-scale feature fusion method, which enables the model to extract and integrate information at different spatial scales, thereby enhancing the perception of multiple targets, especially when facing targets of different sizes and shapes.

[0149] See also Figure 7 , Figure 7 A structural block diagram of a surface corrosion detection system for a large device is provided for an embodiment of the present invention.

[0150] The present invention provides a large-scale device surface corrosion detection system, comprising:

[0151] An acquisition module 701 is used to acquire a corrosion image of a large device to be detected, and input the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, wherein the preset large device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network, and a target detection head;

[0152] The feature extraction module 702 is used to use a fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected, and generate an enhanced feature map;

[0153] The convolution fusion module 703 is used to perform feature convolution fusion on the enhanced feature map through the target neck network and output a fused feature map;

[0154] The target detection module 704 is used to input the fused feature map into the target detection head for target detection and output the corrosion detection result.

[0155] Further, the fusion attention feature extraction network includes a convolution fusion attention mechanism module and a multi-scale feature fusion module; the enhanced feature map includes a first enhanced feature map, a second enhanced feature map and a third enhanced feature map; the feature extraction module 702 includes:

[0156] The first submodule is used to perform convolution fusion on the corrosion image of the large device to be detected by using a convolution fusion attention mechanism module to generate a first convolution fusion feature map;

[0157] The second submodule is used to perform convolution fusion on the first convolution fusion feature map through a convolution fusion attention mechanism module, and output a first enhanced feature map;

[0158] The third submodule is used to input the first enhanced feature map into the convolution fusion attention mechanism module for convolution fusion to determine the second enhanced feature map;

[0159] The fourth submodule is used to perform convolution fusion on the second enhanced feature map by using the convolution fusion attention mechanism module, and output a second convolution fusion feature map;

[0160] The fifth submodule is used to perform multi-scale feature fusion on the second convolution fusion feature map through a multi-scale feature fusion module to generate a third enhanced feature map.

[0161] Optionally, the convolution fusion attention mechanism module includes two serially connected convolution layers, a convolution block attention mechanism layer, a splicing layer, and a convolution layer; the image processing steps of the convolution fusion attention mechanism module include:

[0162] The eroded input map of the input convolution fusion attention mechanism module is convolved through two series-connected convolution layers to generate the first convolution feature map;

[0163] Segment the first convolution feature map, and output a first segmentation feature map and a second segmentation feature map;

[0164] Using two serially connected convolutional layers to perform a convolution operation on the second segmentation feature map to generate a second convolutional feature map;

[0165] The feature weight of the second convolution feature map is adjusted through the convolution block attention mechanism layer, and the attention feature map of the first convolution block is output;

[0166] The attention feature map of the first convolution block and the second segmentation feature map are concatenated to generate a first concatenated feature map, and the first concatenated feature map is input into two series-connected convolutional layers for convolution operation, and a third convolutional feature map is output;

[0167] The convolution block attention mechanism layer is used to adjust the feature weight of the third convolution feature map to generate the second convolution block attention feature map;

[0168] Concatenate the attention feature map of the second convolution block and the first concatenated feature map, and output a second concatenated feature map;

[0169] Splicing the first segmentation feature map, the first convolution block attention feature map, the second splicing feature map, and the second segmentation feature map to generate a third splicing feature map;

[0170] The third concatenated feature map is convolved through a convolution layer to output a corrosion output map.

[0171] Furthermore, the multi-scale feature fusion module includes three cascaded average pooling layers, three cascaded maximum pooling layers, a splicing layer and a convolution layer; the fifth submodule is specifically used for:

[0172] Performing a convolution operation on the second convolution fusion feature map through a convolution layer to generate an enhanced convolution feature map;

[0173] The enhanced convolution feature map is input into three cascaded average pooling layers and three cascaded maximum pooling layers for average pooling and maximum pooling, respectively, and the average pooling feature map and the maximum pooling feature map are output;

[0174] Use the splicing layer to splice the maximum pooling feature map, the average pooling feature map and the enhanced convolution feature map, and output the spliced ​​pooling feature map;

[0175] The convolution operation is performed on the concatenated pooled feature map through the convolution layer to generate a third enhanced feature map.

[0176] Furthermore, the target neck network includes a partial convolution fusion module, a splicing layer, an upsampling layer, and a convolution layer; the fused feature map includes a first fused feature map, a second fused feature map, and a third fused feature map; the convolution fusion module 703 is specifically used for:

[0177] An upsampling layer is used to upsample the third enhanced feature map, and a first upsampled feature map is outputted, and the first upsampled feature map and the second enhanced feature map are spliced ​​through a splicing layer to generate a first intermediate splicing feature map;

[0178] Input the first intermediate splicing feature map into the partial convolution fusion module for partial convolution fusion, output the first partial convolution fusion feature map, and use the upsampling layer to upsample the first partial convolution fusion feature map to generate a second upsampled feature map;

[0179] The second up-sampled feature map and the first enhanced feature map are spliced ​​through the splicing layer to output a second intermediate spliced ​​feature map, the second intermediate spliced ​​feature map is input into the partial convolution fusion module for partial convolution fusion, and the first fused feature map is output;

[0180] A convolution layer is used to perform a convolution operation on the first fused feature map to generate a first fused convolution feature map, and the first fused convolution feature map and the first partial convolution fused feature map are input into a splicing layer for splicing, and a third intermediate splicing feature map is output;

[0181] Performing partial convolution fusion on the third intermediate spliced ​​feature map through a partial convolution fusion module to generate a second fused feature map, and performing a convolution operation on the second fused feature map using a convolution layer to output a second fused convolution feature map;

[0182] The second fused convolution feature map and the third enhanced feature map are input into the splicing layer for splicing, and a fourth intermediate splicing feature map is output. The fourth intermediate splicing feature map is partially convolutionally fused through a partial convolution fusion module to generate a third fused feature map.

[0183] In an optional embodiment, the system further includes:

[0184] The first module is used to obtain a corrosion image dataset of a large-scale device to be trained, and annotate the corrosion image dataset of the large-scale device to be trained to generate an annotated corrosion image dataset;

[0185] The second module is used to combine the labeled corrosion image data set using a preset mosaic method to generate multiple combined corrosion images;

[0186] The third module is used to crop and adjust each combined erosion image to generate multiple adjusted erosion images;

[0187] The fourth module is used to construct a target corrosion image data set using multiple adjusted corrosion images;

[0188] The fifth module is used to use the target corrosion image data set to train the initial large-scale device surface corrosion detection model and determine the trained preset large-scale device surface corrosion detection model.

[0189] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, modules and sub-modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0190] An embodiment of the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory; when the computer program is executed by the processor, the processor executes the steps of the large-scale device surface corrosion detection method as described in any of the above embodiments.

[0191] An embodiment of the present invention further provides a computer-readable storage medium having a computer program / instruction stored thereon. When the computer program / instruction is executed by a processor, the steps of the large-scale device surface corrosion detection method as in any of the above embodiments are implemented.

[0192] An embodiment of the present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implements the steps of the large-scale device surface corrosion detection method as described in any of the above embodiments.

[0193] In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0194] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0195] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting surface corrosion of a large device, characterized in that: include: Acquire a corrosion image of a large device to be detected, and input the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, wherein the preset large device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network, and a target detection head; Using the fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected, and generate an enhanced feature map; Performing feature convolution fusion on the enhanced feature map through the target neck network, and outputting a fused feature map; The fused feature map is input into the target detection head for target detection, and the corrosion detection result is output.

2. The large-scale device surface corrosion detection method according to claim 1 is characterized in that: The fusion attention feature extraction network includes a convolution fusion attention mechanism module and a multi-scale feature fusion module; the enhanced feature map includes a first enhanced feature map, a second enhanced feature map and a third enhanced feature map; the fusion attention feature extraction network is used to perform feature enhancement extraction on the corrosion image of the large device to be detected to generate an enhanced feature map, including: Using a convolution fusion attention mechanism module to perform convolution fusion on the corrosion image of the large device to be detected to generate a first convolution fusion feature map; Performing convolution fusion on the first convolution fusion feature map through a convolution fusion attention mechanism module, and outputting a first enhanced feature map; Inputting the first enhanced feature map into a convolutional fusion attention mechanism module for convolutional fusion to determine a second enhanced feature map; Using a convolution fusion attention mechanism module to perform convolution fusion on the second enhanced feature map, and output a second convolution fusion feature map; The second convolution fusion feature map is subjected to multi-scale feature fusion through a multi-scale feature fusion module to generate a third enhanced feature map.

3. The large-scale device surface corrosion detection method according to claim 2 is characterized in that: The convolution fusion attention mechanism module includes two serially connected convolution layers, a convolution block attention mechanism layer, a splicing layer, and a convolution layer; the image processing steps of the convolution fusion attention mechanism module include: The eroded input map of the input convolution fusion attention mechanism module is convolved through two series-connected convolution layers to generate the first convolution feature map; Segmenting the first convolution feature map, and outputting a first segmentation feature map and a second segmentation feature map; Using two convolutional layers connected in series to perform a convolution operation on the second segmentation feature map to generate a second convolutional feature map; Adjusting the feature weight of the second convolution feature map through the convolution block attention mechanism layer, and outputting the first convolution block attention feature map; Splicing the first convolution block attention feature map and the second segmentation feature map to generate a first spliced ​​feature map, and inputting the first spliced ​​feature map into two serially connected convolutional layers for convolution operation, and outputting a third convolution feature map; Using a convolutional block attention mechanism layer to adjust the feature weight of the third convolutional feature map to generate a second convolutional block attention feature map; Splicing the second convolution block attention feature map and the first splicing feature map, and outputting a second splicing feature map; Splicing the first segmentation feature map, the first convolution block attention feature map, the second splicing feature map, and the second segmentation feature map to generate a third splicing feature map; A convolution operation is performed on the third concatenated feature map through a convolution layer to output a corrosion output map.

4. The large-scale device surface corrosion detection method according to claim 2, characterized in that: The multi-scale feature fusion module includes three cascaded average pooling layers, three cascaded maximum pooling layers, a splicing layer and a convolution layer; the multi-scale feature fusion module performs multi-scale feature fusion on the second convolution fusion feature map to generate a third enhanced feature map, including: Performing a convolution operation on the second convolution fusion feature map through a convolution layer to generate an enhanced convolution feature map; Inputting the enhanced convolutional feature map into three cascaded average pooling layers and three cascaded maximum pooling layers for average pooling and maximum pooling, respectively, and outputting an average pooling feature map and a maximum pooling feature map; Using a splicing layer to splice the maximum pooling feature map, the average pooling feature map and the enhanced convolution feature map, and output a spliced ​​pooling feature map; A convolution operation is performed on the concatenated pooled feature map through a convolution layer to generate a third enhanced feature map.

5. The large-scale device surface corrosion detection method according to claim 2, characterized in that: The target neck network includes a partial convolution fusion module, a splicing layer, an upsampling layer, and a convolution layer; the fused feature map includes a first fused feature map, a second fused feature map, and a third fused feature map; the enhanced feature map is subjected to feature convolution fusion through the target neck network to output a fused feature map, including: Upsampling the third enhanced feature map by using an upsampling layer to output a first upsampling feature map, and splicing the first upsampling feature map and the second enhanced feature map by using a splicing layer to generate a first intermediate splicing feature map; Inputting the first intermediate splicing feature map into the partial convolution fusion module for partial convolution fusion, outputting a first partial convolution fusion feature map, and upsampling the first partial convolution fusion feature map using an upsampling layer to generate a second upsampled feature map; Splicing the second up-sampled feature map and the first enhanced feature map through a splicing layer, outputting a second intermediate splicing feature map, inputting the second intermediate splicing feature map into a partial convolution fusion module for partial convolution fusion, and outputting a first fused feature map; Using a convolution layer to perform a convolution operation on the first fused feature map to generate a first fused convolution feature map, and inputting the first fused convolution feature map and the first partial convolution fused feature map into a splicing layer for splicing, and outputting a third intermediate splicing feature map; Performing partial convolution fusion on the third intermediate spliced ​​feature map through a partial convolution fusion module to generate a second fused feature map, and performing a convolution operation on the second fused feature map using a convolution layer to output a second fused convolution feature map; The second fused convolution feature map and the third enhanced feature map are input into the splicing layer for splicing, a fourth intermediate splicing feature map is output, and the fourth intermediate splicing feature map is partially convolutionally fused through a partial convolution fusion module to generate a third fused feature map.

6. The large-scale device surface corrosion detection method according to claim 1, characterized in that: The training process of the preset large-scale device surface corrosion detection model includes: Acquire a corrosion image dataset of a large-scale device to be trained, and annotate the corrosion image dataset of the large-scale device to be trained to generate an annotated corrosion image dataset; Using a preset mosaic method to combine the labeled corrosion image data set to generate a plurality of combined corrosion images; Performing cropping and adjustment on each of the combined eroded images respectively to generate a plurality of adjusted eroded images; Using a plurality of the adjusted corrosion images to construct a target corrosion image data set; The target corrosion image data set is used to perform model training on an initial large-scale device surface corrosion detection model, and the trained preset large-scale device surface corrosion detection model is determined.

7. A large-scale device surface corrosion detection system, characterized in that: include: An acquisition module, used for acquiring a corrosion image of a large device to be detected, and inputting the corrosion image of the large device to be detected into a preset large device surface corrosion detection model, wherein the preset large device surface corrosion detection model includes a fusion attention feature extraction network, a target neck network and a target detection head; A feature extraction module, used to use the fusion attention feature extraction network to perform feature enhancement extraction on the corrosion image of the large device to be detected, and generate an enhanced feature map; A convolution fusion module, used for performing feature convolution fusion on the enhanced feature map through the target neck network, and outputting a fused feature map; The target detection module is used to input the fused feature map into the target detection head for target detection and output the corrosion detection result.

8. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large-scale device surface corrosion detection method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the large-scale device surface corrosion detection method as described in any one of claims 1-6 is implemented.

10. A computer program product, characterized in that The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer executes the large-scale device surface corrosion detection method as described in any one of claims 1-6.

Images