Visual building mold corrosion identification method

Abstract

The invention relates to a visual building mold corrosion identification method, which comprises the following steps of: firstly, acquiring a steel mold surface image, inputting the steel mold surface image into an improved YOLOv8 network, and outputting the category and the position of a corrosion area after the steel mold surface image is sequentially processed by a trunk network, a neck network and a head network; the improvement point comprises the following steps: replacing an original C2f module with an SE-C2f module in a backbone network, and introducing a channel attention mechanism to enhance small target feature expression; the neck network adopts a BiFPN bidirectional feature pyramid structure, and light-weight cross-scale information interaction is realized by adding transverse jump connection pointing to low resolution from high resolution and learning weight rapid normalization fusion; and the head network adopts an ASFF-Head module to perform up-sampling or down-sampling on different layers of feature maps and then fuse the feature maps, so that the spatial scale invariance is improved and feature conflicts are inhibited. The method can be widely applied to an intelligent cleaning and maintenance system of the steel mold while keeping real-time performance.

Description

Technical Field

[0001] This invention relates to the field of industrial visual analysis, specifically a visual method for identifying rust on building molds. Background Technology

[0002] In the field of industrial automation, especially in construction, steel molds are key tools for concrete forming, and their surface quality directly affects the construction quality and the performance of the final concrete product. However, during long-term use, steel molds are prone to corrosion and impurity accumulation due to exposure to harsh natural environments and frequent mechanical friction. This corrosion and impurities not only affect the appearance of the mold, but more importantly, they significantly reduce the adhesion between the mold and the concrete, leading to defects on the concrete surface such as holes and cracks, which in turn affect the safety and durability of the overall structure.

[0003] Currently, most cleaning methods for steel mold surfaces still rely on a comprehensive cleaning strategy, meaning that regardless of whether rust or impurities are present on the mold surface, it is cleaned indiscriminately. This method is particularly inefficient when dealing with localized rust, as large areas without rust are also included in the cleaning process, resulting in unnecessary water and energy consumption. Furthermore, traditional manual or basic mechanical cleaning methods are not only inefficient but also struggle to ensure consistent cleaning quality, easily leaving cleaning blind spots and leading to rust residue.

[0004] Traditional cleaning methods often use chemical cleaning agents, which can release harmful gases during use, polluting the environment and posing a threat to the health of operators. Furthermore, manual cleaning requires a large workforce, is labor-intensive, and its effectiveness is significantly affected by human factors.

[0005] With the development of computer vision technology, object detection algorithms have been widely used in industrial inspection. However, existing object detection algorithms, especially for the detection of small objects, still have significant shortcomings. For example, the YOLO (YouOnly Look Once) series of algorithms, as an efficient object detection framework, performs excellently in real-time performance, but its accuracy often fails to meet the high standards of industrial inspection when detecting small-sized rust defects.

[0006] Although researchers have made improvements to the YOLO algorithm in recent years to address its shortcomings in small target detection, such as by introducing GA-RPN and multi-task FPN structures, embedding attention mechanisms and optimizing path aggregation networks, and integrating improved CBAM attention mechanisms, these methods have improved detection performance to some extent. However, problems such as slow processing speed, insufficient real-time performance, or high model complexity still exist, making it difficult to achieve efficient and accurate rust detection in industrial settings. Summary of the Invention

[0007] The present invention aims to overcome the shortcomings of the prior art and provide a visual method for identifying corrosion of building molds, thereby solving the problem of detecting corrosion and impurities on the surface of steel molds.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: A visual method for identifying corrosion of architectural molds, characterized in that it includes: Obtain images of the steel mold surface; The image is input into an improved YOLOv8 network model, which processes it sequentially through the trunk, neck, and head networks to output the category and location information of the rusted area. In this context, the C2f module in the backbone network is replaced by the SE-C2f module, which embeds an SE-Bottleneck structure in the residual branch for channel attention weighting. The neck network adopts a BiFPN bidirectional feature pyramid structure to achieve bidirectional fusion of multi-scale features from top to bottom and from bottom to top. The head network uses the ASFF-Head module to perform adaptive spatial feature fusion on feature maps of different scales to complete the detection.

[0009] The aforementioned visual method for identifying corrosion of building molds is characterized in that: the SE-C2f module introduces a new channel in the residual connection, which consists of a global pooling layer, a fully connected layer, and a Sigmoid activation function, to improve data processing efficiency and detection accuracy.

[0010] The aforementioned visual method for identifying corrosion of building molds is characterized by the following: the BiFPN structure adds a lateral jump connection from the high-resolution feature map to the low-resolution feature map on the basis of the original YOLOv8 PANet, forming a bidirectional fusion path that is parallel from top to bottom and from bottom to top; multi-source input features at the same scale layer are rapidly normalized and weighted through learnable weights to reduce the number of parameters and achieve cross-scale information interaction.

[0011] The aforementioned visual method for identifying corrosion of building molds is characterized in that: for each detection scale, the ASFF-Head module first unifies the feature maps from different network layers to that detection scale through upsampling or downsampling, and then performs pixel-wise weighted fusion using a spatial weight map generated by 1×1 convolution; the spatial weight map is normalized by Softmax and multiplied channel-wise with the multi-source feature map to obtain fused features, which are used for subsequent classification and regression branches.

[0012] The aforementioned visual method for identifying corrosion of architectural molds is characterized in that the SE-Bottleneck structure comprises, in sequence: The first 1×1 convolution is used for dimensionality reduction; 3×3 depthwise convolution is used for spatial feature extraction; The SE weight calculation unit consists of global average pooling, fully connected compression, ReLU activation, fully connected recovery, and Sigmoid activation, and outputs channel weights which are multiplied channel by channel with the input features. The second 1×1 convolution is used for dimensionality increase; And residual connections, used to add the weighted features to the original input.

[0013] The aforementioned visual method for identifying rust on architectural molds is characterized in that: the improved YOLOv8 network model employs joint supervision of CIoU loss and Focal Loss during the training phase, with CIoU loss used for bounding box regression and Focal Loss used to alleviate the imbalance between positive and negative samples for small rust targets. The aforementioned visual method for identifying corrosion of building molds is characterized in that: before image input, the surface of the steel mold is photographed in a linear strip pattern, and adjacent images retain a 10% overlap area; By registering overlapping areas through template matching or feature point matching, complete surface splicing with a width of 2.5m can be achieved. The stitched image is slidably cropped into sub-images according to a preset window. The size of the sub-images is consistent with the input size of the improved YOLOv8 network model. They are then sequentially fed into the network to complete the corrosion detection, and the detection results are mapped back to the original stitched coordinate system.

[0014] The aforementioned visual method for identifying corrosion of building molds is characterized by: calculating the proportion of corrosion area and the maximum corrosion diameter based on the detected corrosion area outline; When the percentage of rusted area exceeds the first threshold or the maximum rust diameter exceeds the second threshold, it is marked as an area that needs mechanical grinding. The coordinates of the center and corner points of the rusted area are converted to the equipment's world coordinate system and output to the downstream intelligent cleaning device to guide the sandblasting or grinding head to operate as needed.

[0015] The aforementioned visual method for identifying corrosion of building molds is characterized in that: the improved YOLOv8 network model is deployed on an embedded GPU or industrial PC, the overall number of parameters does not exceed 1.2 times that of YOLOv8n, the detection speed is ≥45FPS, the mAP@0.5 is improved by ≥3.5% compared to the original YOLOv8n, and the mAP@0.5:0.95 is improved by ≥3.3%.

[0016] The beneficial effects of this invention are as follows: As can be seen from the above technical solution, this application provides a visual method for identifying rust on construction molds, using the YOLOv8 algorithm to detect the cleaning area on the surface of the steel mold. Addressing the complexity of the construction environment, the randomness of image quality, and the problems of slow processing speed and insufficient real-time performance due to the large data volume, this invention improves and optimizes the YOLOv8 algorithm, mainly involving the SE-C2f module, BiFPN structure, and ASFF detection head.

[0017] This invention improves and optimizes the YOLOv8 algorithm to enhance the detection accuracy of rusted areas in steel molds. The main improvements based on the YOLOv8 model are threefold: First, the C2f module in YOLOv8 is improved by replacing the original Bottleneck module with the SEBottleneck module, forming the SE-C2f module; second, the neck network of YOLOv8 is replaced with a BiFPN bidirectional feature pyramid network; and third, the ASFF-Head module is used to optimize the head network of YOLOv8.

[0018] Experimental results show that improvements to the YOLOv8 model not only enhance detection accuracy but also reduce missed detections of small targets. While the YOLOv8n+BiFPN+SE-C2f+ASFF-Detect model resulted in a 28.63 frames per second decrease in FPS, it achieved improvements of 3.64% and 3.44% respectively in the key metrics of mAP@0.5 and mAP@0.5:0.95 compared to the original YOLOv8n model. This indicates that the introduction of BiFPN effectively reduced model complexity, achieving lightweight design, while the combination of SE-C2f and ASFF-Detect provided optimal performance for the detection task. Attached Figure Description

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments: Figure 1 Diagrams of SE-C2f and SEBottleneck structures; Figure 2 Here is a diagram of the BiFPN structure; Figure 3 This is a structural diagram of the ASFF detection head; Figure 4 Functional view of the mold cleaning visual inspection system. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection claimed in this application. like Figure 2-4 As shown: A visual method for identifying rust on architectural molds, comprising: Obtain images of the steel mold surface; The image is input into an improved YOLOv8 network model, which processes it sequentially through the trunk, neck, and head networks to output the category and location information of the rusted area. In this context, the C2f module in the backbone network is replaced by the SE-C2f module, which embeds an SE-Bottleneck structure in the residual branch for channel attention weighting. The neck network adopts a BiFPN bidirectional feature pyramid structure to achieve bidirectional fusion of multi-scale features from top to bottom and from bottom to top. The head network uses the ASFF-Head module to perform adaptive spatial feature fusion on feature maps of different scales to complete the detection.

[0021] The SE-C2f module improves data processing efficiency and detection accuracy by introducing a new channel in the residual connections, which consists of a global pooling layer, a fully connected layer, and a sigmoid activation function.

[0022] The BiFPN structure adds a lateral skip connection from the high-resolution feature map to the low-resolution feature map on the basis of the original PANet of YOLOv8, forming a bidirectional fusion path that is parallel from top to bottom and bottom to top. Multi-source input features of the same scale layer are quickly normalized and weighted through learnable weights to reduce the number of parameters and realize cross-scale information interaction.

[0023] For each detection scale, the ASFF-Head module first unifies the feature maps from different network layers to that detection scale by upsampling or downsampling, and then performs pixel-wise weighted fusion using the spatial weight map generated by 1×1 convolution. The spatial weight map is normalized by Softmax and multiplied channel-wise with the multi-source feature map to obtain the fused features, which are used for subsequent classification and regression branches.

[0024] The SE-Bottleneck structure comprises, in sequence: The first 1×1 convolution is used for dimensionality reduction; 3×3 depthwise convolution is used for spatial feature extraction; The SE weight calculation unit consists of global average pooling, fully connected compression, ReLU activation, fully connected recovery, and Sigmoid activation, and outputs channel weights which are multiplied channel by channel with the input features. The second 1×1 convolution is used for dimensionality increase; And residual connections, used to add the weighted features to the original input.

[0025] The improved YOLOv8 network model employs joint supervision of CIoU loss and Focal Loss during the training phase. CIoU loss is used for bounding box regression, while Focal Loss is used to alleviate the imbalance between positive and negative samples in small rust targets. Before image input, the surface of the steel mold is photographed in a linear strip pattern, with adjacent images retaining a 10% overlap area. By registering overlapping areas through template matching or feature point matching, complete surface splicing with a width of 2.5m can be achieved. The stitched image is slidably cropped into sub-images according to a preset window. The size of the sub-images is consistent with the input size of the improved YOLOv8 network model. They are then sequentially fed into the network to complete the corrosion detection, and the detection results are mapped back to the original stitched coordinate system.

[0026] Based on the detected rust area outline, calculate the rust area percentage and the maximum rust diameter; When the percentage of rusted area exceeds the first threshold or the maximum rust diameter exceeds the second threshold, it is marked as an area that needs mechanical grinding. The coordinates of the center and corner points of the rusted area are converted to the equipment's world coordinate system and output to the downstream intelligent cleaning device to guide the sandblasting or grinding head to operate as needed.

[0027] The improved YOLOv8 network model is deployed on embedded GPUs or industrial PCs, with an overall parameter count not exceeding 1.2 times that of YOLOv8n, a detection speed of ≥45 FPS, and an improvement of ≥3.5% in mAP@0.5 compared to the original YOLOv8n, and an improvement of ≥3.3% in mAP@0.5:0.95.

[0028] The above technical solution mainly improves the YOLOv8 algorithm, and is summarized as follows: 1. SE-C2f module This paper improves the C2f module in YOLOv8, forming the SE-C2f module. By introducing an attention mechanism, it enhances the model's ability to identify rust defects and mitigate interference from complex background textures. Figure 1As shown, the SEBottleneck module replaces the original C2f module's Bottleneck module. By introducing a new channel in the residual connection, consisting of a global pooling layer, a fully connected layer, and a Sigmoid activation function, the data processing efficiency and detection accuracy are improved.

[0029] 2. BiFRN structure Replacing the neck network of YOLOv8 with a BiFPN bidirectional feature pyramid network enables a lightweight algorithm design, reducing the number of parameters and computational complexity, and improving real-time performance. The BiFPN structure achieves efficient information flow and fusion between feature maps of different scales through two feature fusion paths, one top-down and one bottom-up, thereby improving detection accuracy. Figure 2 As shown, the BiFPN structure is an optimization of the original neck network of YOLOv8, adding a path from high resolution to low resolution, which is beneficial for feature fusion, achieving a lightweight effect and improving detection accuracy.

[0030] 3. ASFF detection head To address the issue of missed detections of densely packed small targets, the YOLOv8 header network is optimized by employing the ASFF-Head module. For example... Figure 3 As shown, ASFF-Head performs feature fusion by upsampling or downsampling feature maps and learning weight parameters to generate the final fused features used for prediction. ASFF-Head is beneficial for fusing spatial feature information from feature maps of different sizes before classification and regression, adaptively selecting the most useful features, filtering conflicting information, improving spatial scale invariance, enhancing model accuracy, and maintaining detection speed.

[0031] This invention improves and optimizes the YOLOv8 algorithm to enhance the detection accuracy of rusted areas in steel molds. The main improvements based on the YOLOv8 model are threefold: First, the C2f module in YOLOv8 is improved by replacing the original Bottleneck module with the SEBottleneck module, forming the SE-C2f module; second, the neck network of YOLOv8 is replaced with a BiFPN bidirectional feature pyramid network; and third, the ASFF-Head module is used to optimize the head network of YOLOv8.

[0032] Experimental results show that improvements to the YOLOv8 model not only enhance detection accuracy but also reduce missed detections of small targets. While the YOLOv8n+BiFPN+SE-C2f+ASFF-Detect model resulted in a 28.63 frames per second decrease in FPS, it achieved improvements of 3.64% and 3.44% respectively in the key metrics of mAP@0.5 and mAP@0.5:0.95 compared to the original YOLOv8n model. This indicates that the introduction of BiFPN effectively reduced model complexity, achieving lightweight design, while the combination of SE-C2f and ASFF-Detect provided optimal performance for the detection task.

[0033] The control system of the cleaning equipment uses an industrial control computer, a touch screen, and a PLC main control unit as its core control module. This module coordinates the various components of the equipment to complete tasks such as rust removal, rinsing, drying, and applying release agent. The actual required grinding width is 2.5m, with a recognition accuracy of ±1mm. Based on the actual situation and cost-effectiveness, the vision system uses two Huarui A3B00CG000 industrial cameras with a resolution of 5472×3648. Each camera has a field of view of 1.3m, and the overlap between the two cameras is 10cm, which perfectly meets the 2.5m grinding width requirement. The pixel accuracy is μ = 1300mm / 5472 ≈ 0.24mm / pixel, meeting the accuracy requirements. The length direction is achieved through multiple feeds and step-by-step photographing for grinding.

[0034] The visual inspection system is divided into five sub-modules (such as...) Figure 4 As shown in the figure, the modules are: vision hardware module, communication module, vision inspection module, human-computer interaction module, and intelligent rust removal module.

[0035] 1. The vision hardware module consists of an industrial computer responsible for data and image processing, and image acquisition tools such as cameras and lenses for capturing images. These hardware components will collaborate with the software system. 2. The communication module covers the communication interfaces between devices and the interactive communication links with the vision inspection application. 3. The vision inspection module can complete the camera calibration process before inspection, establish a unified camera coordinate system, and realize the conversion from pixel coordinates to device world coordinates. During inspection, it can read the configuration file containing mold information, perform surface inspection, identify rusted areas and cleaning boundary corners, and perform coordinate transformation. Finally, it outputs the inspection results in a predefined format. 4. The human-machine interaction module needs to set and receive rust removal process parameters and mold configuration information, read inspection data, and feed the information back to the system. 5. The intelligent rust removal module combines the recognition results provided by the vision inspection module, uses image processing technology to evaluate the degree of rust, and finally adjusts the process parameter configuration to complete the intelligent rust removal process of the equipment.

[0036] The above are merely embodiments provided in this application and are not intended to limit this application. Although this application has been described in detail with reference to the embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. However, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A visual method for identifying corrosion of architectural molds, characterized in that... It includes: Obtain images of the steel mold surface; The image is input into an improved YOLOv8 network model, which processes it sequentially through the trunk, neck, and head networks to output the category and location information of the rusted area. In this context, the C2f module in the backbone network is replaced by the SE-C2f module, which embeds an SE-Bottleneck structure in the residual branch for channel attention weighting. The neck network adopts a BiFPN bidirectional feature pyramid structure to achieve bidirectional fusion of multi-scale features from top to bottom and from bottom to top. The head network uses the ASFF-Head module to perform adaptive spatial feature fusion on feature maps of different scales to complete the detection.

2. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: The SE-C2f module improves data processing efficiency and detection accuracy by introducing a new channel in the residual connections, which consists of a global pooling layer, a fully connected layer, and a sigmoid activation function.

3. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: The BiFPN structure adds a lateral skip connection from the high-resolution feature map to the low-resolution feature map on the basis of the original PANet of YOLOv8, forming a bidirectional fusion path that is parallel from top to bottom and bottom to top. Multi-source input features of the same scale layer are quickly normalized and weighted through learnable weights to reduce the number of parameters and realize cross-scale information interaction.

4. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: For each detection scale, the ASFF-Head module first unifies the feature maps from different network layers to that detection scale by upsampling or downsampling, and then performs pixel-wise weighted fusion using the spatial weight map generated by 1×1 convolution. The spatial weight map is normalized by Softmax and multiplied channel-wise with the multi-source feature map to obtain the fused features, which are used for subsequent classification and regression branches.

5. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: The SE-Bottleneck structure comprises, in sequence: The first 1×1 convolution is used for dimensionality reduction; 3×3 depthwise convolution is used for spatial feature extraction; The SE weight calculation unit consists of global average pooling, fully connected compression, ReLU activation, fully connected recovery, and Sigmoid activation, and outputs channel weights which are multiplied channel by channel with the input features. The second 1×1 convolution is used for dimensionality increase; And residual connections, used to add the weighted features to the original input.

6. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: The improved YOLOv8 network model employs joint supervision of CIoU loss and Focal Loss during the training phase. CIoU loss is used for bounding box regression, while Focal Loss is used to alleviate the imbalance between positive and negative samples in small rust targets.

7. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: Before image input, the surface of the steel mold is photographed in a linear strip pattern, with adjacent images retaining a 10% overlap area. By registering overlapping areas through template matching or feature point matching, complete surface splicing with a width of 2.5m can be achieved. The stitched image is slidably cropped into sub-images according to a preset window. The size of the sub-images is consistent with the input size of the improved YOLOv8 network model. They are then sequentially fed into the network to complete the corrosion detection, and the detection results are mapped back to the original stitched coordinate system.

8. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: Based on the detected rust area outline, calculate the rust area percentage and the maximum rust diameter; When the percentage of rusted area exceeds the first threshold or the maximum rust diameter exceeds the second threshold, it is marked as an area that needs mechanical grinding. The coordinates of the center and corner points of the rusted area are converted to the equipment's world coordinate system and output to the downstream intelligent cleaning device to guide the sandblasting or grinding head to operate as needed.

9. The visual method for identifying corrosion of architectural molds according to claim 1, characterized in that: The improved YOLOv8 network model is deployed on embedded GPUs or industrial PCs, with an overall parameter count not exceeding 1.2 times that of YOLOv8n, a detection speed of ≥45 FPS, and an improvement of ≥3.5% in mAP@0.5 compared to the original YOLOv8n, and an improvement of ≥3.3% in mAP@0.5:0.95.

Images