Blade defect detection method based on progressive feature enhancement and Inner-IoU

Abstract

The invention discloses a blade defect detection method based on progressive feature enhancement and Inner-IoU, and relates to the technical field of computer vision and image processing, and the method comprises the following steps: constructing and preprocessing a blade defect data set, and adopting a joint data enhancement JDA strategy to expand a training data set; a ResNet50 backbone network of an integrated channel self-adaptive weighted ACAW module is constructed, after the preprocessed image is input into the ResNet50 backbone network, feature maps of different scales are generated, and for the feature map of each scale, the system executes an attention-based channel self-adaptive weighted ACAW processing flow; in the encoding and decoding stage, a progressive feature input PFI encoder is adopted; according to the method, the detection speed is remarkably improved, the calculation amount is greatly reduced through the combination of the PFI strategy and the light-weight feed-forward network LW-FFN, the model can meet the real-time detection requirement of an industrial assembly line, and through the synergistic effect of the ACAW module and the Inner-IoU loss function, it is ensured that tiny defect features are not submerged by the background and can be accurately positioned.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a blade defect detection method based on progressive feature enhancement and Inner-IoU. Background Technology

[0002] The performance of aero-engines directly determines the safety and economy of an aircraft. Turbine blades, as the most critical components in aero-engines operating in the harshest and most stressful environment, are subjected to extreme conditions of high temperature, high pressure, high speed, and highly corrosive airflow for extended periods. During casting and service, blade surfaces are highly susceptible to various defects such as cracks, porosity, pits, scratches, and foreign matter. If these minute defects are not effectively detected before assembly, they can easily propagate into fatigue fracture sources during subsequent high-stress operation, leading to blade breakage or even catastrophic accidents such as in-flight engine shutdown. Therefore, achieving high-precision, automated detection of surface defects in aero-engine blades is of paramount engineering importance for ensuring flight safety and extending engine service life.

[0003] Currently, defect detection technology for aero-engine blades has evolved from manual visual inspection and traditional non-destructive testing to modern computer vision inspection. However, existing technologies still have significant limitations when dealing with complex industrial scenarios.

[0004] (1) Traditional detection techniques and their limitations Early and current blade inspections primarily relied on manual visual inspection (including the use of endoscopes) and traditional non-destructive testing (NDT) techniques, such as eddy current testing, ultrasonic testing, magnetic particle testing, and fluorescent penetrant testing. Manual inspection is highly dependent on the operator's subjective experience and concentration; prolonged repetitive work easily leads to visual fatigue, resulting in a high rate of missed defects, and the testing standards are difficult to quantify and standardize. While traditional NDT methods can detect internal defects, their sensitivity to minute surface defects is limited. For example, magnetic particle testing is easily affected by the blade surface geometry, producing false magnetic traces and causing false alarms; while fluorescent penetrant testing is complex and time-consuming, making it difficult to meet the stringent efficiency requirements of modern production lines. Overall, traditional methods have inherent drawbacks such as low efficiency, high subjectivity, and low automation, making them unsuitable for the intelligent transformation needs of the aerospace manufacturing industry.

[0005] (2) Detection techniques based on convolutional neural networks (CNN) and their limitations With the explosive growth of deep learning technology, object detection algorithms centered on convolutional neural networks (CNNs) have been increasingly applied to industrial surface defect detection. These mainly include two-stage algorithms represented by Faster R-CNN and one-stage algorithms represented by YOLO and SSD. Although these algorithms perform well on general datasets, they have revealed significant structural shortcomings in aero-engine blade inspection tasks. First, CNNs primarily rely on the sliding of convolutional kernels to extract features, resulting in a limited receptive field and difficulty in establishing long-distance pixel dependencies. For common blade surface defects such as slender cracks or large-scale diffuse defects, CNNs often suffer from incomplete feature extraction due to a lack of global contextual information. Second, most existing CNN detectors rely on hand-designed anchor box mechanisms and non-maximum suppression (NMS) post-processing. The size and proportion of the anchor boxes are usually set based on prior knowledge, making it difficult to cover the extremely wide scale range of blade defects (from micrometer-level pores to centimeter-level scratches); while NMS operations are prone to missed detections or false negatives when processing dense or overlapping defects. Furthermore, as the number of network layers increases, although CNNs can extract strong semantic features, the downsampling process leads to a significant loss of location information for minor defects, severely limiting the detection accuracy of small targets.

[0006] (3) Transformer-based (DETR) detection technology and its pain points To overcome the limitations of CNNs, a Transformer-based end-to-end object detector (DETR) was introduced into the field. DETR utilizes a self-attention mechanism to capture global features of the image and achieves a one-to-one correspondence between predicted and ground truth boxes through bipartite graph matching, completely eliminating anchor boxes and Non-Maximum Suppression (NMS). However, directly applying the original DETR architecture to the detection of defects in aero-engine blades faces three irreconcilable technical contradictions.

[0007] The first issue is the conflict between computational complexity and real-time performance. The computational cost of the self-attention mechanism in Transformer increases quadratically with the length of the input feature sequence (i.e., the number of pixels in the feature map). In order to detect minute defects on the blade surface, the input image usually needs to be high-resolution, which results in extremely long feature sequences, exponential increases in computational resource consumption, and extremely slow inference speed, far from meeting the millisecond-level real-time detection standards of industrial production lines.

[0008] Secondly, there is a contradiction between the difficulty in converging on small targets and the inaccurate localization. DETR models converge extremely slowly during training, and in the early stages of training, the attention mechanism often struggles to focus on sparse, small defect features, easily becoming dominated by complex background textures. More critically, existing bounding box regression loss functions (such as IoU or GIoU Loss) are extremely insensitive to scale. For tiny pore defects, even a small pixel difference between the predicted and ground truth bounding boxes can cause significant fluctuations or even zero in the Intersection over Union (IoU), resulting in extremely low gradient backpropagation efficiency and making it difficult for the network to learn accurate localization information.

[0009] Thirdly, there is a contradiction between extreme sample imbalance and feature redundancy. In real-world aerospace manufacturing scenarios, the vast majority of blades are qualified, with very few defective samples, and the number of different defect types (such as common scratches and rare cracks) varies greatly. This severe long-tail distribution can cause the model to overfit to the background or the majority class of defects during training. Simultaneously, the complex textures of blade images (such as metallic reflections and machining marks) mean that simple feature fusion methods will introduce a large amount of background noise into deep networks, masking the already weak defect feature responses and leading to a decline in detection performance.

[0010] In summary, existing detection technologies cannot simultaneously meet the comprehensive requirements of high accuracy, high speed, and resistance to sample imbalance. Therefore, developing a novel detection method that combines global perception capabilities, possesses a lightweight inference architecture, and is specifically optimized for small targets and imbalanced samples has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0011] The purpose of this invention is to provide a blade defect detection method based on progressive feature enhancement and Inner-IoU, so as to solve the problems existing in the prior art mentioned in the background art.

[0012] To achieve the above objectives, the present invention provides the following technical solution: The blade defect detection method based on progressive feature enhancement and Inner-IoU includes the following steps: S1: Construct and preprocess the blade defect dataset, and use the Joint Data Augmentation (JDA) strategy to expand the training dataset; S2: Construct a ResNet50 backbone network with an integrated channel adaptive weighted ACAW module. After inputting the preprocessed image into the ResNet50 backbone network, feature maps of different scales are generated. For each feature map, the system executes an attention-based channel adaptive weighted ACAW processing flow. S3: In the encoding and decoding stages, a progressive feature input PFI encoder is used, and the decoder receives feature representations and object queries from the encoder. Each block in the decoder adopts a lightweight feedforward network LW-FFN structure. S4: Calculate the Inner-IoU loss and backpropagate. During the training iteration, the system calculates the Inner-IoU loss between the predicted bounding box and the true bounding box. The system weights and sums the regression loss and classification loss, and updates the network parameters through the backpropagation algorithm until the model converges. In the inference stage, the model directly outputs the category of the defect and its precise spatial coordinates, thereby completing the task of detecting defects in aero-engine blades.

[0013] Preferably, the specific steps of S1 are as follows: The system acquires raw image data of aero-engine blades. During the training phase, it performs Joint Data Augmentation (JDA) operations. The system randomly selects K images as references and performs random scaling, cropping, and color jittering on each image. Mosaic enhancement is performed by stitching four randomly selected images into a single large image and recalculating the corresponding label coordinates. Based on this, a Mixup operation is triggered with a certain probability, which linearly superimposes two images and their labels according to a randomly generated scaling factor λ to generate an enhanced training set containing rich background information and multi-scale targets. This enhanced training set is then input into the network for training.

[0014] Preferably, the specific steps of S2 are as follows: Feature maps are processed using Global Average Pooling (GAP). Compression is performed in the spatial dimension to obtain the channel descriptor. The descriptor is then input into a bottleneck structure, Bottleneck, which contains two fully connected layers. The first layer reduces the number of channels to reduce the number of parameters. After passing through the ReLU activation function, the second layer restores the number of channels by increasing their dimensionality. Normalized weight vectors are generated using the Sigmoid function. and the weight vector Compared with the original feature map Perform channel-by-channel multiplication to obtain the enhanced feature map. .

[0015] Preferably, the specific steps of using the progressive feature input PFI encoder in S3 are as follows: In the first to N layers of the Transformer encoder, the first M layers only receive low-level feature maps, the middle layers gradually introduce feature maps that fuse mid-level semantic information, and the high-level feature maps only participate in self-attention calculation in the deep layers of the encoder. The feature maps at each scale are first unified in channel dimension through a 1x1 convolution operation, and then gradually input into the Transformer encoder to participate in self-attention operations in a hierarchical order from shallow to deep.

[0016] Preferably, the specific steps of LW-FFN-based decoding and non-shared prediction in S3 are as follows: The input tensor is split into two parts along the channel dimension. One part is directly retained, and the other part is input into a concave multilayer perceptron (MLP) for processing. The concave structure reduces the number of neurons in the intermediate hidden layer by first reducing the dimension and then increasing the dimension, thereby reducing computational complexity. The two features are spliced ​​together and a channel shuffling operation is performed to enhance the information exchange capability between different channels; At the prediction output end, the system constructs two independent detection heads, one for defect category classification and the other for defect bounding box regression. The two do not share any weight parameters and calculate the classification loss and regression loss respectively.

[0017] Preferably, the processing procedure of S4 is as follows: Define scale factor Based on the scale factor Calculate the center point coordinates of the ground truth bounding box and the predicted bounding box respectively. and generate a width of Height is Given the inner auxiliary boxes, calculate the intersection-union ratio (CUC) between the two auxiliary boxes, denoted as . This serves as the basis for determining the loss.

[0018] Preferably, the regression loss function is defined as .

[0019] Preferably, the scale factor The value range is set to 0.7-0.8.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention significantly improves the detection speed. By combining the PFI strategy with LW-FFN, the computational load is greatly reduced, enabling the model to meet the real-time detection requirements of industrial production lines.

[0021] 2. This invention improves the detection rate of minute defects. The synergistic effect of the ACAW module and the Inner-IoU loss function ensures that the features of minute defects are not submerged by the background and can be accurately located.

[0022] 3. This invention enhances the robustness of the model, and the joint data augmentation strategy effectively overcomes the problems of sample scarcity and imbalance in industrial scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the method of the present invention.

[0025] Figure 3 This is a comparison chart of the detection results of the method of the present invention and existing methods. Detailed Implementation

[0026] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0027] Please see Figure 1-3 The present invention provides the following technical solutions: The specific implementation process of the method proposed in this invention relies on a computer vision processing system, which includes image acquisition equipment (such as an industrial X-ray machine or a high-resolution camera) and a computing unit equipped with a high-performance GPU. The specific steps of the implementation of this invention are described in detail below.

[0028] The blade defect detection method based on progressive feature enhancement and Inner-IoU includes the following steps: S1: Construct and preprocess the blade defect dataset, and use the Joint Data Augmentation (JDA) strategy to expand the training dataset.

[0029] The system acquires raw image data of aero-engine blades. During the training phase, it performs Joint Data Augmentation (JDA) operations. The system randomly selects K images as references and performs random scaling, cropping, and color jittering on each image. Then, Mosaic enhancement is performed, which stitches four randomly selected images into a large image and recalculates the corresponding label coordinates. Based on this, a Mixup operation is triggered with a certain probability, which linearly superimposes two images and their labels according to a randomly generated scaling factor λ to generate an enhanced training set containing rich background information and multi-scale targets, and inputs it into the network for training.

[0030] In the data preprocessing stage, this invention employs a Joint Data Augmentation (JDA) strategy to augment the training dataset. This strategy does not rely solely on traditional geometric transformations but organically combines three techniques: Mixup, Mosaic, and Mix-Mos. Specifically, Mixup enhances the model's ability to linearly model different defect features by linearly interpolating and mixing two images and their labels; Mosaic enriches the background context of the images and significantly increases the number of small targets in a single input image by randomly cropping and stitching four images; Mix-Mos further introduces a mixing mechanism based on Mosaic. This joint strategy effectively simulates the complex and varied background of industrial sites and, without increasing inference costs, solves the problem of extreme imbalance between positive and negative samples at the data level, thereby improving the model's generalization robustness.

[0031] S2: Construct a ResNet50 backbone network integrating channel adaptive weighted ACAW modules. After inputting the preprocessed image into the ResNet50 backbone network, extract feature maps at three different scales, corresponding to the outputs of different layers of the backbone network, denoted as C3, C4, and C5 respectively. For each scale feature map... The system executes an attention-based channel adaptive weighting (ACAW) processing flow.

[0032] Feature maps are processed using Global Average Pooling (GAP). Compression is performed in the spatial dimension to obtain the channel descriptor. The descriptor is then input into a bottleneck structure, Bottleneck, which contains two fully connected layers. The first layer reduces the number of channels to reduce the number of parameters. After passing through the ReLU activation function, the second layer restores the number of channels by increasing their dimensionality. Normalized weight vectors are generated using the Sigmoid function. and the weight vector Compared with the original feature map Perform channel-by-channel multiplication to obtain the enhanced feature map. This process mathematically reweights key defect features, effectively suppressing the interference of irrelevant background textures on the detection results.

[0033] In the feature extraction stage, this invention constructs a backbone network with an Integrated Channel Adaptive Weighted (ACAW) module. After the original image is input into the ResNet50 backbone network, feature maps of different scales are generated. Unlike spatial attention mechanisms (such as coordinate attention) commonly used in existing technologies, the ACAW module focuses on feature recalibration in the channel dimension. Given that blade defects typically have high response values ​​only on specific feature channels, ACAW first performs global average pooling on the input feature map to obtain global contextual information, and then captures the nonlinear dependencies between channels through a multilayer perceptron (MLP) to generate channel weight vectors. This weight vector is multiplied by the original feature map, thereby adaptively suppressing background noise channels and significantly enhancing channel features containing key defect information such as cracks and pores, providing high-quality feature input for subsequent processing.

[0034] S3: In the encoding and decoding stages, a progressive feature input PFI encoder is used. The decoder receives feature representations and object queries from the encoder. Each block in the decoder adopts a lightweight feed-forward network (LW-FFN) structure.

[0035] In the encoding stage, instead of flattening and stitching feature maps C3, C4, and C5 together and inputting them into the Transformer encoder at once, the system uses Progressive Feature Input (PFI) instead.

[0036] In the first to N layers of the Transformer encoder, the first M layers only receive low-level feature maps (such as C3 after ACAW processing), the intermediate layers gradually introduce feature maps that fuse mid-level semantic information, and the high-level feature maps only participate in self-attention calculation in the deep layers of the encoder. Feature maps at each scale are first unified in channel dimension through a 1x1 convolution operation, and then progressively input into the Transformer encoder in a progressive order of depth to participate in self-attention operations. Through this progressive feature input mechanism, the number of tokens processed by the shallow encoder is significantly reduced, thereby effectively reducing memory usage at the hardware level and improving the model's inference frame rate (FPS).

[0037] The input tensor is split into two parts along the channel dimension. One part is retained directly, while the other part is input into a concave multilayer perceptron (MLP) for processing. The concave structure reduces the number of neurons in the intermediate hidden layer by first reducing the dimension and then increasing the dimension, thereby reducing computational complexity. The two parts of features are then concatenated, and a channel shuffle operation is performed to enhance the information interaction capability between different channels.

[0038] At the prediction output end, the system constructs two independent detection heads, one for defect category classification and the other for defect bounding box regression. The two do not share any weight parameters and calculate the classification loss and regression loss respectively.

[0039] In the encoding and decoding stages, this invention designs a Progressive Feature Input (PFI) encoder and a Lightweight Feedforward Network (LW-FFN). Addressing the high computational complexity of traditional Transformers, the PFI strategy changes the input method of feature maps. Instead of flattening the input all at once, the multi-scale feature maps enhanced by ACAW are progressively input into different layers of the Transformer encoder according to semantic hierarchy. This design significantly reduces the number of tokens in the self-attention mechanism, lowering the computational complexity from quadratic and significantly improving inference speed. Simultaneously, the feedforward network in the decoder is replaced with an LW-FFN structure. This structure utilizes channel splitting, channel shuffling, and a concave MLP design to significantly compress the number of parameters while maintaining feature transformation capabilities. Furthermore, the prediction head adopts a non-shared parameter design, where the classification and regression branches are optimized independently to compensate for any potential accuracy loss due to the lightweight design.

[0040] S4: Calculate the Inner-IoU loss and backpropagate. During training iterations, the system calculates the predicted bounding boxes. With the true bounding box Inner-IoU loss between them. First, define the scaling factor. , scale factor The value range is set to 0.7-0.8. Based on the aforementioned scale factor... Calculate the center point coordinates of the ground truth bounding box and the predicted bounding box respectively. and generate a width of Height is Given the inner auxiliary boxes, calculate the intersection-union ratio (CUC) between the two auxiliary boxes, denoted as . This serves as the basis for determining the loss.

[0041] The regression loss function is defined as The system performs a weighted sum of regression and classification losses and updates network parameters through backpropagation until the model converges. During the inference phase, the model directly outputs the defect category and its precise spatial coordinates, thus completing the defect detection task for aero-engine blades.

[0042] During the model training and optimization phases, this invention introduces the Inner-IoU loss function to replace the traditional IoU loss. For the micro-pores and fine scratches commonly found in aero-engine blades, traditional IoU suffers from extremely low gradient backpropagation efficiency when the overlap between the predicted and ground truth boxes is low, leading to slow convergence. The Inner-IoU loss function introduces a scale factor to generate smaller auxiliary boxes within both the target and predicted boxes, and calculates the IoU loss based on these auxiliary boxes. This method allows the model to focus more on the overlap of the core target region, effectively amplifying gradient changes caused by small deviations, thereby accelerating the model's convergence speed on small targets and significantly improving localization accuracy.

[0043] Two-stage detection algorithms, single-stage CNN algorithms, YOLO series algorithms, Transformer-type algorithms, and the method of this invention were tested under the same conditions. The performance of each algorithm is shown in Table 1.

[0044] Table 1 Performance Comparison of Various Algorithms As shown in Table 1, under the same dataset, same training / test partition, same input size, same confidence threshold / IoU threshold, and same training epochs / learning rate strategy, the method of this invention (improved DETR) achieves the best results in precision (P), recall (R), and mAP@0.5, which are 0.823, 0.758, and 0.771, respectively. Its overall detection performance is superior to two-stage detection algorithms, single-stage CNN algorithms, YOLO series algorithms, and existing Transformer-type algorithms. Specifically, compared to the original DETR (P=0.742, R=0.689, mAP@0.5=0.704), the method of this invention improves by 0.081, 0.069, and 0.067, respectively, indicating that this invention has significant improvements in defect localization accuracy and detection completeness. Compared to the representative Transformer-type method Deformable... The DETR (mAP@0.5=0.726) and RT-DETR (mAP@0.5=0.710) values ​​of the proposed method are improved by 0.045 and 0.061 respectively, indicating that the proposed method effectively enhances the characterization ability of complex backgrounds and small defects while taking into account the advantages of the end-to-end detection framework. Compared with YOLOv12 (P=0.807, R=0.746, mAP@0.5=0.752), which has better performance in the YOLO series, the proposed method still improves by 0.016, 0.012 and 0.019 respectively, further verifying the comprehensive advantages and stability of the proposed method in the task of detecting defects in aero-engine blades. In terms of model lightweighting and real-time performance, as shown in Table 1, the parameter count of the proposed method is only 34.2M, which is reduced by 50.6% and 45.1% compared with Cascade R-CNN (69.2M) and the latest YOLOv12 (62.3M). This is mainly due to the introduction of the LW-FFN structure in the decoder, which significantly compresses redundant parameters. In terms of inference speed, the method of this invention achieves 56.8 FPS, far exceeding Faster R-CNN (18.6 FPS) and the original DETR (14.8 FPS), meeting the real-time detection standard of over 30 FPS for industrial production lines. Although slightly lower than the pure convolutional architecture of the YOLO series, considering the significant advantage of this invention in mAP accuracy (0.771), this inference speed achieves the best balance between accuracy and efficiency.

[0045] Ablation experiments were conducted on the method of the present invention, and the results of the ablation experiments are shown in Table 2:

[0046] Table 2 Ablation Experiment Results Table 2 shows that the performance of the baseline model (A) built based on the original DETR is P=0.742, R=0.689, and mAP@0.5=0.704. After introducing Joint Data Augmentation (JDA) on the baseline (B), the three metrics improved to 0.763, 0.712, and 0.726, respectively, representing improvements of 0.021, 0.023, and 0.022 compared to the baseline. This indicates that JDA can effectively alleviate the problem of insufficient generalization caused by sample scarcity and scene changes. Further adding the ACAW module (C) further improves the metrics to 0.789, 0.731, and 0.749, representing improvements of 0.026, 0.019, and 0.023 compared to B, indicating that the channel self-... Adaptive weighting can highlight key defect information and suppress background noise interference. Further introduction of PFI and LW-FFN structures (D) improves the performance metrics to 0.807, 0.746, and 0.762, respectively, representing improvements of 0.018, 0.015, and 0.013 compared to C. This demonstrates that progressive feature input and lightweight feedforward networks can maintain and improve detection accuracy while reducing computational overhead. Adding the Inner-IoU loss function to D forms the final model (E), achieving performance metrics of 0.823, 0.758, and 0.771, respectively, representing improvements of 0.016, 0.012, and 0.009 compared to D. This indicates that Inner-IoU can further enhance the learning efficiency of bounding box regression and improve the localization accuracy of small defects. The improvement from Model C to Model D, thanks to the introduction of the PFI progressive strategy and the LW-FFN structure, achieved a leap in inference speed from 13.5 FPS to 56.5 FPS, fully demonstrating the key role of this lightweight design in eliminating the computational bottleneck of the Transformer. A comprehensive comparison from A to E shows that each improved module has a cumulative gain in performance improvement, verifying the effectiveness and necessity of the "JDA+ACAW+PFI / LW-FFN+Inner-IoU" multi-strategy collaborative design described in this invention.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A blade defect detection method based on progressive feature enhancement and Inner-IoU, characterized in that, Includes the following steps: S1: Construct and preprocess the blade defect dataset, and use the Joint Data Augmentation (JDA) strategy to expand the training dataset; S2: Construct a ResNet50 backbone network with an integrated channel adaptive weighted ACAW module. After inputting the preprocessed image into the ResNet50 backbone network, feature maps of different scales are generated. For each feature map, the system executes an attention-based channel adaptive weighted ACAW processing flow. S3: In the encoding and decoding stages, a progressive feature input PFI encoder is used, and the decoder receives feature representations and object queries from the encoder. Each block in the decoder adopts a lightweight feedforward network LW-FFN structure. S4: Calculate the Inner-IoU loss and backpropagate. During the training iteration, the system calculates the Inner-IoU loss between the predicted bounding box and the true bounding box. The system weights and sums the regression loss and classification loss, and updates the network parameters through the backpropagation algorithm until the model converges. In the inference stage, the model directly outputs the category of the defect and its precise spatial coordinates, thereby completing the task of detecting defects in aero-engine blades.

2. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 1, characterized in that, The specific steps of S1 are as follows: The system acquires raw image data of aero-engine blades. During the training phase, it performs Joint Data Augmentation (JDA) operations. The system randomly selects K images as references and performs random scaling, cropping, and color jittering on each image. Mosaic enhancement is performed by stitching four randomly selected images into a single large image and recalculating the corresponding label coordinates. Based on this, a Mixup operation is triggered with a certain probability, which linearly superimposes two images and their labels according to a randomly generated scaling factor λ to generate an enhanced training set containing rich background information and multi-scale targets. This enhanced training set is then input into the network for training.

3. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 1, characterized in that, The specific steps of S2 are as follows: Feature maps are processed using Global Average Pooling (GAP). Compression is performed in the spatial dimension to obtain the channel descriptor. The descriptor is then input into a bottleneck structure, Bottleneck, which contains two fully connected layers. The first layer reduces the number of channels to reduce the number of parameters. After passing through the ReLU activation function, the second layer restores the number of channels by increasing their dimensionality. Normalized weight vectors are generated using the Sigmoid function. and the weight vector Compared with the original feature map Perform channel-by-channel multiplication to obtain the enhanced feature map. .

4. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 1, characterized in that, The specific steps of using the progressive feature input PFI encoder in S3 are as follows: In the first to N layers of the Transformer encoder, the first M layers only receive low-level feature maps, the middle layers gradually introduce feature maps that fuse mid-level semantic information, and the high-level feature maps only participate in self-attention calculation in the deep layers of the encoder. The feature maps at each scale are first unified in channel dimension through a 1x1 convolution operation, and then gradually input into the Transformer encoder to participate in self-attention operations in a hierarchical order from shallow to deep.

5. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 1, characterized in that, The specific steps of LW-FFN-based decoding and non-shared prediction in S3 are as follows: The input tensor is split into two parts along the channel dimension. One part is directly retained, and the other part is input into a concave multilayer perceptron (MLP) for processing. The concave structure reduces the number of neurons in the intermediate hidden layer by first reducing the dimension and then increasing the dimension, thereby reducing computational complexity. The two features are spliced ​​together and a channel shuffling operation is performed to enhance the information exchange capability between different channels; At the prediction output end, the system constructs two independent detection heads, one for defect category classification and the other for defect bounding box regression. The two do not share any weight parameters and calculate the classification loss and regression loss respectively.

6. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 1, characterized in that, The processing procedure of S4 is as follows: Define scale factor Based on the scale factor Calculate the center point coordinates of the ground truth bounding box and the predicted bounding box respectively. and generate a width of Height is Given the inner auxiliary boxes, calculate the intersection-union ratio (CUC) between the two auxiliary boxes, denoted as . This serves as the basis for determining the loss.

7. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 6, characterized in that, The regression loss function is defined as .

8. The blade defect detection method based on progressive feature enhancement and Inner-IoU according to claim 6, characterized in that, The scale factor The value range is set to 0.7-0.8.

Images