Engine blade surface defect identification method and system based on YOLO network and detection device

Abstract

The invention provides an engine blade surface defect identification method and system based on a YOLO network and a detection device. The method comprises the following steps: acquiring a to-be-identified image based on an engine blade; and inputting the to-be-identified image into a pre-trained target detection model to obtain a defect type and a defect bounding box of the to-be-identified image. The invention provides a target detection model which is improved on the basis of a YOLOv8 network to realize surface defect identification of the engine blade, and solves the problems of missing detection, error detection, inaccurate defect frame positioning and the like existing in a common computer vision detection system.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an engine blade surface defect recognition method, system, and detection device based on a YOLO network. Background Art

[0002] Surface defect detection on engine blades is crucial for every aircraft's pre-takeoff engine reliability inspection. With the advancement of artificial intelligence and deep learning, computer vision-based surface defect detection solutions are becoming increasingly mature. Computer vision inspection systems, such as those using the YOLO series of algorithms, are gradually replacing manual inspection for surface defect detection on engine blades.

[0003] Although the YOLO series of algorithms are relatively mature, there are still major problems in the detection of surface defects in aircraft engine blades, such as missed detection, false detection, detected defects being similar to the detection background, and inaccurate positioning. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method and system for identifying engine blade surface defects based on the YOLO network, which solves the problems of missed detection, wrong detection, and inaccurate defect frame positioning when detecting engine blade surface defects through the YOLOv8 network in the existing technology.

[0005] According to an embodiment of the present invention, a first aspect provides a method for identifying engine blade surface defects based on a YOLO network, comprising:

[0006] Acquire an image to be identified based on an engine blade;

[0007] Inputting the image to be identified into a pre-trained object detection model to obtain the defect type and defect bounding box of the image to be identified;

[0008] The target detection model is based on the YOLOv8 network, including a backbone network, a neck network and three groups of detection heads connected in sequence; the two original Conv structures of the YOLOv8 network are replaced by an average pooling downsampling module in the backbone network, and a feature selection module is inserted; the original C2f module of the YOLOv8 network is replaced by a C2f_Star_CAA module in the neck network; the three groups of detection heads are all dynamic detection heads based on the multi-scale attention mechanism.

[0009] Optionally, the backbone network includes 1st to 11th layer network basic blocks connected in sequence, wherein the 6th layer network basic block and the 8th layer network basic block are the average pooling downsampling modules, and the 11th layer network basic block is the feature selection module;

[0010] The neck network includes network basic blocks from the 12th to the 23rd layers connected in sequence, wherein the network basic block of the 23rd layer is a C2f_Star_CAA module;

[0011] The dynamic detection head includes a first detection head, a second detection head, and a third detection head;

[0012] Among them, the feature selection module is connected to the 10th layer network basic block, the 12th layer network basic block and the C2f_Star_CAA module, the 5th layer network basic block is connected to the 16th layer network basic block, the 7th layer network basic block is connected to the 13th layer network basic block, the first detection head is connected to the 17th layer network basic block and the 18th layer network basic block, the second detection head is connected to the 20th layer network basic block, and the third detection head is connected to the 23rd layer network basic block.

[0013] Optionally, the feature selection module includes a channel attention processing branch and a spatial attention processing branch;

[0014] The feature selection module also fuses the output of the channel attention processing branch and the output of the spatial attention processing branch based on a dual attention mechanism.

[0015] Optionally, the feature selection module implements the following steps:

[0016] S1. Obtain target features and save them as residual connections to obtain target features of the residual connections; the target features are provided by the network basic blocks before the feature selection module;

[0017] S21. On the channel attention processing branch, based on the target features of the residual connection, perform average pooling in the width direction and the height direction to obtain a width direction feature map and a height direction feature map;

[0018] S22, extract features from the width direction feature map and the height direction feature map through 1×3 convolution and 3×1 convolution respectively, and generate a convolved feature map;

[0019] S23, performing transposition processing and matrix multiplication calculation on the convolved feature map, and outputting it after channel expansion;

[0020] S24, processes the output of S23 through 1x1 convolution and activates it through Sigmoid activation function to generate coordinate attention weights;

[0021] S31. On the spatial attention processing branch, performing channel aggregation based on the target features of the residual connection;

[0022] S32, use 7×7 convolution to process the output of S31 to generate spatial weights;

[0023] S4. Based on the coordinate attention weight and the spatial weight, a product fusion operation is performed to obtain a target attention weight. The target attention weight is processed by weight compression and residual weighting. Based on the processed target attention weight, the target feature is extracted and output to the next layer of network basic block.

[0024] Optionally, before inputting the image to be identified into a pre-trained object detection model based on a YOLO network, the method further includes:

[0025] Set the parameters of the C2f_Star_CAA module, including at least the convolution kernel space size, expansion factor, and grouped convolution parameters.

[0026] Optionally, before inputting the image to be identified into a pre-trained object detection model based on a YOLO network, the method further includes:

[0027] The parameters of the dynamic detection head are set, including at least the number of hidden layer channels, the number of stacking times, and the upper limit of the discrete interval.

[0028] Optionally, the dynamic detection head outputs N defect frames to be selected of different scales;

[0029] After the image to be identified is input into the pre-trained target detection model, weighted averaging processing is performed on N defect frames to be selected to obtain the defect boundary frame, where N is a positive integer.

[0030] The second aspect provides an engine blade surface defect recognition system based on a YOLO network, comprising:

[0031] A preprocessing and data enhancement module is used to perform data enhancement on the engine blade defect image obtained by the borescope equipment when pre-training the target detection model; and is also used to pre-process the image to be identified of the engine blade after obtaining the image to be identified;

[0032] A deep learning detection module includes a pre-trained target detection model for obtaining the defect type and defect bounding box of the image to be identified; the target detection model is based on the YOLOv8 network, and includes a backbone network, a neck network, and three groups of detection heads connected in sequence; the backbone network replaces the two original Conv structures of the YOLOv8 network with an average pooling downsampling module, and inserts a feature selection module; the neck network replaces the original C2f module of the YOLOv8 network with a C2f_Star_CAA module; the three groups of detection heads are all dynamic detection heads based on a multi-scale attention mechanism.

[0033] Optionally, a post-processing and result fusion module is also included;

[0034] In the deep learning detection module, the dynamic detection head outputs N defect frames to be selected at different scales;

[0035] The post-processing and result fusion module is used to perform weighted averaging processing on the N defect frames to be selected to obtain the defect boundary frame, where N is a positive integer.

[0036] The third aspect provides a detection device, including an image acquisition device and a processor, wherein the image acquisition device is used to acquire an image to be identified based on an engine blade, and the processor includes the above-mentioned engine blade surface defect recognition system based on the YOLO network, and executes the above-mentioned engine blade surface defect recognition method based on the YOLO network.

[0037] Compared with the prior art, the present invention has the following beneficial effects: providing an improved target detection model based on the YOLOv8 network to realize the recognition of engine blade surface defects. For the original YOLOv8 network, the improvements include: replacing the original two Conv structures with the average pooling downsampling module in the backbone network, and inserting a feature selection module, wherein the average pooling downsampling module is used to generate feature maps of different resolutions to obtain target features of different scales in the image and improve the recognition ability of defects of different sizes, and the feature selection module is used to improve the ability to capture key features, that is, the detection ability to further extract defect features of the identified defects; and also includes replacing an original C2f module with the C2f_Star_CAA module in the neck network, wherein the C2f_Star_CAA module combines C2f (Cross Stage Partial Fractional, cross-stage partial block connection) and Star-CAA (Star-Shaped Channel Attention Aggregation (star-shaped channel attention aggregation) can enhance the key features of the target features, that is, enhance the defect features in the image while retaining details. At the same time, it adopts depth-wise separable convolution and 1×1 convolution reorganization to effectively detect small targets while being lightweight, avoiding false detections and missed detections. Finally, the original detection head is modified to a dynamic detection head based on the multi-scale attention mechanism, which effectively improves detection performance in complex scenarios and overcomes the problem of inaccurate defect bounding box positioning caused by the similarity between defects in the image and the background. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the implementation flow of the engine blade surface defect recognition method based on the YOLO network according to an embodiment of the present invention;

[0039] Figure 2 It is a structural diagram of the existing YOLOv8 network;

[0040] Figure 3 Schematic diagram of the structure of the target detection module according to an embodiment of the present invention;

[0041] Figure 4 Schematic diagram of the structure of the average pooling downsampling module according to an embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram of the structure of a feature selection module according to an embodiment of the present invention;

[0043] Figure 6 This is a schematic diagram of the structure of the C2f_Star_CAA module according to an embodiment of the present invention;

[0044] Figure 7 This is a schematic structural diagram of a dynamic detection head according to an embodiment of the present invention;

[0045] Figure 8 Schematic diagram of the structure of an engine blade surface defect recognition system based on a YOLO network according to an embodiment of the present invention;

[0046] Figure 9 Schematic diagram of the structure of a detection device according to an embodiment of the present invention;

[0047] Figure 10 This is a comparison chart of the recognition results of the engine blade surface defect recognition method based on the existing YOLOv8 network and the YOLO network based on an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0049] like Figure 1 As shown, the embodiment of the present invention proposes a method for identifying engine blade surface defects based on the YOLO network, including but not limited to the following steps:

[0050] S100: Acquire an image to be recognized based on an engine blade.

[0051] S200: Input the image to be identified into a pre-trained object detection model to obtain the defect type and defect bounding box of the image to be identified.

[0052] In a specific application, different positions of the engine blade are photographed, so that multiple images to be identified can be obtained in the above step S100, and then the defects at different positions of the engine blade are obtained through the above step S200, and are represented by the defect type and defect boundary box of the corresponding image to be identified.

[0053] In one embodiment, before the above step S100, the image to be identified based on the engine blade is normalized by the preprocessing and data enhancement module shown below.

[0054] In step S200 above, the present embodiment provides an improved object detection model based on the YOLOv8 network, i.e., a pre-trained object detection model. Improvements include: in the backbone network, the two Conv structures of the YOLOv8 network are replaced with an average pooling downsampling module, and a feature selection module is inserted; in the neck network, the C2f module of the YOLOv8 network is replaced with a C2f_Star_CAA module; and all three detection heads are dynamic detection heads based on a multi-scale attention mechanism. Among them, the average pooling downsampling module is used to generate feature maps of different resolutions to obtain target features of different scales in the image and improve the recognition ability of defects of different sizes; the feature selection module is used to improve the ability to capture key features, that is, the detection ability to further extract defect features from the identified defects; the C2f_Star_CAA module combines C2f (Cross Stage Partial Fractional) and Star-CAA (Star-Shaped Channel Attention Aggregation), which can enhance the key features in the target features, that is, enhance the defect features in the image while retaining details. At the same time, it uses depthwise separable convolution and 1×1 convolution reorganization to effectively detect small targets while being lightweight, avoiding false detection and missed detection; and the dynamic detection head based on the multi-scale attention mechanism can effectively improve detection performance in complex scenarios and overcome the problem of inaccurate defect bounding box positioning caused by the similarity between defects in the image and the background.

[0055] like Figure 2 and Figure 3 As shown, in another embodiment of the present invention, a schematic diagram of the structure of the existing YOLOv8 network and a schematic diagram of the structure of the target detection module in step S200 are respectively given for comparison. Figure 2 The structural diagram of the existing YOLOv8 network shown in Figure 1, and Figure 3 The schematic diagram of the structure of the target detection module in the above step S200 is shown. First, the target detection model in the embodiment of the present invention and the existing YOLOv8 network both include a backbone network, a neck network and three sets of detection heads connected in sequence. Secondly, for the backbone network part, as shown in FIG. Figure 2 and Figure 3 Backbone in both includes 10 layers of network basic blocks, but Figure 3In the target detection model of the embodiment of the present invention, the 6th layer network basic block and the 8th layer network basic block are constructed as an average pooling downsampling module ADown, and Figure 2 In the example, the network basic blocks at the corresponding positions are different in Conv structure; and, in the embodiment of the present invention, the feature selection module ICA is inserted between the 10th-layer network basic block SPPF and the 11th-layer network basic block UPsample, then the feature selection module ICA is the 11th-layer network basic block. In addition, it should be noted that the average pooling downsampling module ADown and the feature selection module ICA can output target features of different scales in the recognition of surface defects of engine blades, and capture independent components such as shape, texture, and color in the target features as key features to improve detection robustness. Then, for the neck network part, such as Figure 2 and Figure 3 Neck in both include 12 layers of network basic blocks, but Figure 3 In the target detection model of the embodiment of the present invention, the 23rd layer network basic block is constructed as a C2f_Star_CAA module, and Figure 2 The network basic block at the corresponding position in is the C2f module. In addition, it should be noted that compared with the C2f module, the C2f_Star_CAA module improves the expression ability through dual-branch nonlinear interaction on the basis of the C2f module, that is, aggregating multi-branch features and combining channel attention mechanism to enhance the key features of the target features, that is, enhance the defect features in the image and retain the details. Finally, for the detection head part, such as Figure 2 and Figure 3 The Head in each includes three sets of detection heads, but Figure 3 In the target detection model of the embodiment of the present invention, the detection head used is a dynamic detection head, namely DyHead_Detect. Among them, DyHead_Detect improves the feature expression ability through dynamic attention and is suitable for complex scenes.

[0056] like Figure 3 As shown, the embodiment of the present invention also modifies some connection relationships, wherein the 11th layer network basic block,

[0057] (i.e., the feature selection module CAA) is connected to the 10th-layer network basic block, the 12th-layer network basic block and the 23rd-layer network basic block (i.e., the C2f_Star_CAA module), the 5th-layer network basic block is connected to the 16th-layer network basic block, the 7th-layer network basic block is connected to the 13th-layer network basic block, the first detection head DyHead_Detect1 is connected to the 17th-layer network basic block and the 18th-layer network basic block, the second detection head DyHead_Detect2 is connected to the 20th-layer network basic block, and the third detection head DyHead_Detect3 is connected to the 23rd-layer network basic block.

[0058] like Figures 4 to 7 As shown, in another embodiment of the present invention, the detailed structure and working principle of the average pooling downsampling module, feature selection module, C2f_Star_CAA module and dynamic detection head used in the above step S200 are also described. Figure 3 The structure of the target detection module shown in FIG5 illustrates the detailed structure and working principle of the average pooling downsampling module, feature selection module, C2f_Star_CAA module, and dynamic detection head.

[0059] In practical applications, the feature selection module captures key features based on the output of the first to tenth layer network basic blocks in the backbone network. Figure 3 The network structure of the target detection model shown in the figure captures the key features, which are the outputs of the 1st to 10th network basic blocks, that is, the target features in the embodiment of the present invention, which mainly include the average pooling downsampling modules of the 6th and 8th network basic blocks, and the feature maps of different resolutions generated.

[0060] Therefore, first of all, Figure 4 As shown in Figure 2, the structure and working principle of the average pooling downsampling module are as follows: It first receives an input tensor x, provided by the 6th or 8th layer network basic block, and is represented as H×W×C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels. It then performs average pooling on the input tensor x, such as Avg pool2d, to obtain (H-1)×(W-1)×C. This is then split into two parts, (H-1)×(W-1)×C / 2. One part is convolved with the half containing Conv3×3 to obtain H / 2×W / 2×C / 2, and the other part is max-pooled with the half containing MaxPool2d and Conv1×1 to obtain H / 2×W / 2×C / 2. Convolution is then applied to each pooled tensor, and the results are concatenated and returned to obtain the target features H / 2×W / 2×C. This process reduces feature loss.

[0061] Then, by Figure 5 The feature selection module shown captures the independent components of the above target features such as shape, texture, and color as key features. Figure 5 The feature selection module consists of a channel attention processing branch 51 and a spatial attention processing branch 52. The feature selection module also fuses the outputs of the channel attention processing branch and the spatial attention processing branch based on a dual attention mechanism. The feature selection module captures the independent components of the target features such as shape, texture, and color as key features. The implementation steps include:

[0062] S1. Obtain target features and save them as residual connections to obtain target features of the residual connections; the target features are provided by the network basic blocks before the feature selection module;

[0063] S21. On the channel attention processing branch, based on the target features of the residual connection, perform average pooling in the width direction and the height direction to obtain a width direction feature map and a height direction feature map;

[0064] S22, extract features from the width direction feature map and the height direction feature map through 1×3 convolution and 3×1 convolution respectively, and generate a convolved feature map;

[0065] S23, transpose the convolved feature map and perform matrix multiplication, and output it after channel expansion; S24, process the output of S23 by 1x1 convolution and activate it by Sigmoid activation function to generate coordinate attention weight;

[0066] S31. In the spatial attention processing branch, channel aggregation is performed based on the target features of the residual connection;

[0067] S32, using 7×7 convolution to process the output of S31 to generate spatial weights;

[0068] S4. Based on the coordinate attention weight and the spatial weight, a product fusion operation is performed to obtain a target attention weight. The target attention weight is processed by weight compression and residual weighting. Based on the processed target attention weight, the target feature is extracted and output to the next layer of network basic block.

[0069] It should be noted that in the above step S1, the target feature is provided by the network basic block before the feature selection module, based on Figure 2 , that is, the outputs of the 1st to 10th layer network basic blocks, especially the average pooling downsampling modules of the 6th and 8th layer network basic blocks, generate feature maps of different resolutions.

[0070] According to the above steps S21 to S24, the actual application process of the channel attention processing branch 51 is the same as Figure 5 The left branch shown corresponds to the following. In actual applications, it first performs mean pooling in the width direction, such as X Avg pool. The input x is averaged along the width dimension (dim = 3) to obtain [B, C, H, 1]. Then, a 1×3 convolution is performed to extract features in the width direction and reduce the dimension to [B, C / / r, H, 1]. Then, the dimension is adjusted and the permute operation is performed to obtain [B, H, C / / r]. The height direction processing is similar to the width direction processing. For example, in Y Avg pool, the input x is averaged along the height dimension (dim = 2) to obtain [B, C, 1, W]. Then, a 3×1 convolution is performed to extract features in the width direction and reduce the dimension to [B, C / / r, 1, W]. Then, the dimension is adjusted and the permute operation is performed to obtain [B, W, C / / r]. Then perform matrix multiplication, such as MatMul, to obtain [B, H, W], and then expand the channel, such as Unsqueeze, to obtain [B, C / / r, H, W]. At the end of this branch, the channel is repeated, and the coordinate attention weights [B, C, H, W] are generated through 1x1 convolution and Sigmoid activation function, such as Conv+Sigmoid. According to the above steps S31 to S32, the spatial attention processing branch 52 is connected with Figure 5 The right branch shown corresponds to, in actual applications, it first performs channel aggregation, performs mean pooling along the channel dimension, such as Avg pool, to generate avg[B,1,H,W], and then performs maximum pooling along the channel dimension, such as Max pool, to generate max[B,1,H,W], and then performs feature splicing, such as Cat, to splice the mean and maximum feature maps into [B,2,H,W], and through convolution and Sigmoid activation functions, such as Conv+Sigmoid, to generate spatial attention weights [B,1,H,W], and finally performs feature reweighting, such as Re-weight, to expand the spatial attention weights to [B,C,H,W]. According to step S4 above, the output of the channel attention processing branch and the output of the spatial attention processing branch are fused based on the dual attention mechanism, including: first, performing a product operation to highlight the common important areas and generating attention weights [B, C, H, W]; then, using SIgmoid activation to compress the weights to [0, 1]; finally, performing residual enhancement, using dynamic residual scaling to avoid over-suppression or under-enhancement problems caused by fixed weights. Where H is the height of the feature map, W is the width of the feature map, C is the number of channels of the feature map, r is the reduction ratio, and W is the batch size, that is, the number of images input to the model at one time.

[0071] like Figure 6As shown in FIG, the detailed structure of the C2f_Star_CAA module is shown. In an embodiment of the present invention, the C2f_Star_CAA module combines C2f (Cross Stage Partial Fractional) and Star-CAA (Star-Shaped Channel Attention Aggregation), which can enhance key features, that is, enhance defect features in the image.

[0072] like Figure 7 As shown in the figure, the detailed structure of the dynamic detection head is shown. In the embodiment of the present invention, the dynamic detection head is mainly used to combine scale-aware attention, spatial attention, and channel attention to complete the object classification task ObjectClassification, center point regression task CenterRegression, and detection box regression task BoxRegression to improve the feature expression capability and is suitable for complex scenes.

[0073] It should be noted that to further improve or balance the detection efficiency and accuracy of surface defects on aircraft engine blades, the parameters of the aforementioned C2f_Star_CAA module and dynamic detection head must be set before the image is input into the pre-trained YOLO network-based object detection model. Exemplarily, the C2f_Star_CAA module parameters are set to at least include the convolution kernel spatial size, expansion factor, and grouped convolution parameters. In a preferred implementation, the convolution kernel spatial size is 7×7 to expand the receptive field, the expansion factor is 0.7 to increase the number of hidden layer channels to strengthen important feature channels, further enhancing the ability to express intermediate features, the grouped convolution parameter is 2 to further reduce the number of parameters, and the number of repeated Star_Block stacks is 3 to enhance feature expression. The dynamic detection head parameters are set to at least include the number of hidden layer channels, the number of stacks, and the upper limit of the discrete interval. In a preferred implementation, the number of hidden layer channels is 512, the number of stacks is 2, and the upper limit of the discrete interval is 16 to enhance feature expression.

[0074] In another embodiment of the present invention, the dynamic detection head is configured to output N candidate defect frames of different scales. After the image to be identified is input into a pre-trained object detection model, before obtaining the defect bounding box of the image to be identified, a weighted average of the N candidate defect frames is performed to obtain the defect bounding box. This reduces duplicate detection and missed detection, ensuring accurate positioning of the output defect frame.

[0075] like Figure 8 As shown, an embodiment of the present invention also provides an engine blade surface defect recognition system based on a YOLO network, which can train a target detection model based on a YOLO network and use the target detection model based on a YOLO network to recognize engine blade surface defects. It includes:

[0076] The preprocessing and data enhancement module 81 is used to perform data enhancement on the engine blade defect image obtained by the borescope equipment when pre-training the target detection model; and is also used to pre-process the image to be identified of the engine blade after obtaining the image to be identified;

[0077] The deep learning detection module 82 includes a target detection model obtained after pre-training, which is used to obtain the defect type and defect bounding box of the image to be identified; the target detection model is based on the YOLOv8 network, and includes a backbone network, a neck network and three groups of detection heads connected in sequence; the backbone network replaces the two original Conv structures of the YOLOv8 network with the average pooling downsampling module, and inserts a feature selection module; the neck network replaces the original C2f module of the YOLOv8 network with the C2f_Star_CAA module; the three groups of detection heads are all dynamic detection heads based on the multi-scale attention mechanism.

[0078] like Figure 3 As shown, the backbone network includes 1st to 11th layer network basic blocks, wherein the 6th layer network basic block and the 8th layer network basic block are average pooling downsampling modules, and the 11th layer network basic block is the feature selection module, which is connected to the 10th layer network basic block and the 12th layer network basic block; the neck network includes 12th to 23rd layer network basic blocks, wherein the 23rd layer network basic block is a C2f_Star_CAA module; the dynamic detection head includes a first detection head, a second detection head, and a third detection head, wherein the first detection head is connected to the 12th layer network basic block and the 18th layer network basic block, the second detection head is connected to the 20th layer network basic block, and the third detection head is connected to the 23rd layer network basic block.

[0079] In one embodiment, a post-processing and result fusion module may also be included. After the dynamic detection head outputs N defect frames of different scales to be selected, the post-processing and result fusion module may perform weighted averaging on the N defect frames to be selected and output a defect boundary frame, thereby reducing repeated detection and missed detection, and ensuring that the output defect frame is accurately positioned.

[0080] In addition, illustratively, an embodiment of the present invention further provides a training method for pre-training an object detection model, including but not limited to the following steps:

[0081] Acquire defect images of the engine blades through a borescope device, and process the defect images to obtain a data set;

[0082] Using a labeling tool such as labelme, label the defect type of each defect image in the dataset;

[0083] The labeled dataset is divided into a training set and a validation set, and then imported into the YOLOv8 network for training;

[0084] The trained network model is verified and evaluated to obtain the optimal network model as the target detection model based on the YOLO network.

[0085] The defect types include at least one or more of cracks, scratches, damage, ablation, and pits.

[0086] Among them, the trained network model is verified and evaluated, including verification and evaluation based on parameters such as accuracy, precision, recall, F1 score, and average precision.

[0087] like Figure 9 As shown, an embodiment of the present invention further provides a detection device, including an image acquisition device 91 and a processor 92, wherein the image acquisition device 91 is used to acquire an image to be identified based on an engine blade, and the processor 92 includes the above-mentioned engine blade surface defect recognition system based on the YOLO network, and executes the above-mentioned engine blade surface defect recognition method based on the YOLO network.

[0088] like Figure 10 As shown, the embodiment of the present invention also provides a comparison chart of the recognition results based on the existing YOLOv8 network and the recognition results of the engine blade surface defect recognition method based on the YOLO network according to the embodiment of the present invention. Figure 10 middle Figure 10 (a) shows the recognition result based on the existing YOLOv8 network. Figure 10 (b) is the recognition result of the engine blade surface defect recognition method based on the YOLO network according to the embodiment of the present invention. It can be seen that the defect type and defect boundary box are shown, but Figure 10 The defect type and defect bounding box shown in (a) have a confidence level lower than Figure 10 (b) shows the defect type and the confidence of the defect bounding box, and Figure 10 In (a), a defect bounding box of the defect type ablation is omitted.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying engine blade surface defects based on a YOLO network, characterized in that: include: Acquire an image to be identified based on an engine blade; Inputting the image to be identified into a pre-trained object detection model to obtain the defect type and defect bounding box of the image to be identified; The target detection model is based on the YOLOv8 network, including a backbone network, a neck network and three groups of detection heads connected in sequence; the two original Conv structures of the YOLOv8 network are replaced by an average pooling downsampling module in the backbone network, and a feature selection module is inserted; the original C2f module of the YOLOv8 network is replaced by a C2f_Star_CAA module in the neck network; the three groups of detection heads are all dynamic detection heads based on the multi-scale attention mechanism.

2. The engine blade surface defect recognition method based on the YOLO network according to claim 1, characterized in that: The backbone network includes 1st to 11th layer network basic blocks connected in sequence, wherein the 6th layer network basic block and the 8th layer network basic block are the average pooling downsampling modules, and the 11th layer network basic block is the feature selection module; The neck network includes network basic blocks from the 12th to the 23rd layers connected in sequence, wherein the network basic block of the 23rd layer is a C2f_Star_CAA module; The dynamic detection head includes a first detection head, a second detection head, and a third detection head; Among them, the feature selection module is connected to the 10th layer network basic block, the 12th layer network basic block and the C2f_Star_CAA module, the 5th layer network basic block is connected to the 16th layer network basic block, the 7th layer network basic block is connected to the 13th layer network basic block, the first detection head is connected to the 17th layer network basic block and the 18th layer network basic block, the second detection head is connected to the 20th layer network basic block, and the third detection head is connected to the 23rd layer network basic block.

3. The engine blade surface defect recognition method based on the YOLO network according to claim 1 or 2, characterized in that: The feature selection module includes a channel attention processing branch and a spatial attention processing branch; The feature selection module also fuses the output of the channel attention processing branch and the output of the spatial attention processing branch based on a dual attention mechanism.

4. The engine blade surface defect recognition method based on the YOLO network according to claim 3, characterized in that: The feature selection module implements the following steps: S1. Obtain target features and save them as residual connections to obtain target features of the residual connections; the target features are provided by the network basic blocks before the feature selection module; S21. On the channel attention processing branch, based on the target features of the residual connection, perform average pooling in the width direction and the height direction to obtain a width direction feature map and a height direction feature map; S22, extract features from the width direction feature map and the height direction feature map through 1×3 convolution and 3×1 convolution respectively, and generate a convolved feature map; S23, performing transposition processing and matrix multiplication calculation on the convolved feature map, and outputting it after channel expansion; S24, processes the output of S23 through 1x1 convolution and activates it through the Sigmoid activation function to generate coordinate attention weights; S31. On the spatial attention processing branch, performing channel aggregation based on the target features of the residual connection; S32, use 7×7 convolution to process the output of S31 to generate spatial weights; S4. Based on the coordinate attention weight and the spatial weight, a product fusion operation is performed to obtain a target attention weight. The target attention weight is processed by weight compression and residual weighting. Based on the processed target attention weight, the target feature is extracted and output to the next layer of network basic block.

5. The engine blade surface defect recognition method based on the YOLO network according to claim 4, characterized in that: Before inputting the image to be identified into the pre-trained object detection model based on the YOLO network, the following steps are included: Set the parameters of the C2f_Star_CAA module, including at least the convolution kernel space size, expansion factor, and grouped convolution parameters.

6. The engine blade surface defect recognition method based on the YOLO network according to claim 4, characterized in that: Before inputting the image to be identified into the pre-trained object detection model based on the YOLO network, the following steps are included: The parameters of the dynamic detection head are set, including at least the number of hidden layer channels, the number of stacking times, and the upper limit of the discrete interval.

7. The engine blade surface defect recognition method based on the YOLO network according to claim 1, characterized in that: The dynamic detection head outputs N defect frames to be selected at different scales; After the image to be identified is input into the pre-trained target detection model, weighted averaging processing is performed on N defect frames to be selected to obtain the defect boundary frame, where N is a positive integer.

8. An engine blade surface defect recognition system based on YOLO network, characterized in that: include: The preprocessing and data enhancement module is used to perform data enhancement on engine blade defect images obtained by borescope equipment when pre-training the target detection model; It is also used to pre-process the image to be identified after acquiring the image to be identified of the engine blade; A deep learning detection module includes a pre-trained target detection model for obtaining the defect type and defect bounding box of the image to be identified; the target detection model is based on the YOLOv8 network, and includes a backbone network, a neck network, and three groups of detection heads connected in sequence; the backbone network replaces the two original Conv structures of the YOLOv8 network with an average pooling downsampling module, and inserts a feature selection module; the neck network replaces the original C2f module of the YOLOv8 network with a C2f_Star_CAA module; the three groups of detection heads are all dynamic detection heads based on a multi-scale attention mechanism.

9. The engine blade surface defect recognition system based on the YOLO network according to claim 8, characterized in that: It also includes post-processing and result fusion modules; In the deep learning detection module, the dynamic detection head outputs N defect frames to be selected at different scales; The post-processing and result fusion module is used to perform weighted averaging processing on the N defect frames to be selected to obtain the defect boundary frame, where N is a positive integer.

10. A detection device, characterized in that: The invention comprises an image acquisition device and a processor, wherein the image acquisition device is used to acquire an image to be identified based on an engine blade, the processor comprises the engine blade surface defect recognition system based on the YOLO network as described in claim 8 or 9, and executes the engine blade surface defect recognition method based on the YOLO network as described in any one of claims 1 to 7.

Images