Aero-engine blade defect detection method based on improved YOLOV8 model

Abstract

The invention discloses an aero-engine blade defect detection method based on an improved YOLOV8 model, and belongs to the technical field, in the aero-engine blade defect detection method based on the improved YOLOV8 model, by replacing a backbone network with a bidirectional feature pyramid structure, the model can more efficiently fuse feature information of different scales, effectively capture deep features of micro defects such as micron-sized cracks and air holes, and improve the detection accuracy. And the problem of missing detection caused by tiny defect forms or random distribution is reduced. Meanwhile, the introduced attention mechanism inhibits the influence of interference factors such as blade surface reflection and greasy dirt on the detection result by dynamically adjusting the feature weight, so that the recognition of the defect area under the complex background is more accurate, and the recognition accuracy of the defect area is improved by optimizing the multi-scale feature fusion structure and simplifying redundant calculation nodes. The model keeps high precision, meanwhile, the computing resource consumption is remarkably reduced, and it is ensured that the detection speed meets the real-time requirement of an industrial scene. The design of the decoupling detection head enables the classification and positioning tasks to be processed independently, and the parallel processing capability of the model for multiple types of defects is improved.

Description

Technical Field

[0001] This invention belongs to the field of defect identification and detection technology, specifically a defect detection method for aero-engine blades based on an improved YOLOV8 model. Background Technology

[0002] As a core component, the manufacturing quality of aero-engine blades directly impacts flight safety and equipment lifespan. Traditional inspection methods rely on manual visual inspection or mechanical scanning equipment, which suffer from low efficiency, high subjectivity, and high false negative rates. With the development of industrial automation, machine vision-based inspection technology is gradually becoming more widespread. However, blade defects (such as cracks, corrosion, and porosity) in complex backgrounds remain difficult to accurately identify due to their small size, random distribution, and low contrast. Especially in the aerospace field, the high reflectivity and complex curved surface structure of blade materials (such as titanium alloys and nickel-based superalloys) further increase imaging noise and inspection difficulty.

[0003] However, the current traditional YOLOv8 has significant shortcomings in the aerospace blade scenario: 1. Limited ability to extract deep features of defects such as micron-level cracks and pores, resulting in a high false negative rate; 2. Reflective surfaces and oil stains on the blade surface can easily lead to false detections; 3. Imbalance between computational efficiency and accuracy. Furthermore, samples of aeronautical defects are scarce and of poor imaging quality. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting defects in aero-engine blades based on an improved YOLOV8 model in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: a method for detecting defects in aero-engine blades based on an improved YOLOv8 model, the method comprising the following steps. S1: Collect borehole images of aero-engines as raw data, and use annotation tools to annotate the bounding boxes of defects such as cracks, corrosion, and porosity. Divide the data into training set, validation set and test set, and perform data augmentation by rotation, scaling, translation, inversion and contrast adjustment to expand the sample diversity.

[0006] S2: The original backbone network is replaced with a bidirectional feature pyramid network (BiFPN). By optimizing multi-scale feature fusion through cross-scale connections and learnable weights, the ability to extract deep features of small defects is enhanced, while reducing redundant nodes to reduce computational cost.

[0007] S3: Add a convolutional block attention module (CBAM) before the C2f module of the Backbone network. Dynamically adjust feature weights through channel attention and spatial attention branches to suppress surface reflection and oil pollution interference, and improve the saliency of defect spatial features.

[0008] S4: Embed SE-Net into the C2f module of the Neck network, compress the channel dimension of the feature map through global pooling, and recalibrate the channel weights using fully connected layers and activation functions to enhance the information transmission of key defect channels and improve the generalization of the model.

[0009] S5: BiFPN is used to replace the traditional PANet structure. It fuses feature maps of different resolutions through bidirectional cross-layer connections and combines a weighted feature pyramid to enhance the detection effect of small target defects and reduce the consumption of computing resources.

[0010] S6: Replace the original CIoU loss function with the EIoU loss function. By minimizing the distance between the center point of the predicted box and the true box and penalizing the difference in width and height, the positioning accuracy of irregular defects such as cracks and pores is improved.

[0011] S7: Design three independent detection blocks in the Head network, corresponding to 80×80, 40×40 and 20×20 resolutions respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The loss is output through independent convolutional layers to improve detection efficiency.

[0012] S8: Scale the preprocessed image to 640×640×3 and normalize it. Load the improved YOLOV8 model for parameter initialization. Use an iterative training strategy. After each training round, use the validation set to evaluate the precision and recall until the model converges.

[0013] S9: Load the optimal parameter model to perform inference on the test set, output the defect category, confidence level and bounding box coordinates, combine the borehole images to generate visualized detection results, and calculate the false negative rate and false positive rate to verify the practicality of the model in complex industrial scenarios.

[0014] In a preferred embodiment, in step S1, borehole images of aero-engines are acquired as raw data, covering typical defect types such as cracks, corrosion, and porosity. Annotation tools are used to accurately label the defect regions in each image with bounding boxes, ensuring a consistent annotation format. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Data augmentation employs rotation angles of ±30 degrees, scaling ratios of 0.8 to 1.2 times, horizontal and vertical translation ranges of ±15%, mirror reversal, and contrast adjustment ranges of 0.5 to 1.5 times. Diverse samples are generated through random combinations of augmentation strategies to improve the model's robustness to noise and complex backgrounds.

[0015] In a preferred embodiment, in step S2, the original backbone network is replaced with a bidirectional feature pyramid network (BiFPN). Feature maps of different resolutions are fused through four layers of cross-scale connections, with learnable weight coefficients introduced in each layer to dynamically adjust feature importance. The input image size is fixed at 640×640×3, and the BiFPN output feature map sizes are 80×80, 40×40, and 20×20, respectively. Removing redundant nodes reduces computation by 18%, while retaining cross-layer skip connections to enhance the efficiency of small defect feature transfer.

[0016] In a preferred embodiment, in step S3, a convolutional block attention module (CBAM) is embedded before the C2f module of the backbone network. The channel attention branch generates channel weights through global average pooling and a fully connected layer, while the spatial attention branch generates a spatial weight matrix using a 7×7 convolutional kernel. After the feature map undergoes dual-branch weighting, the channel dimension is compressed to 1 / 16 of its original size, and the spatial weights are normalized using the sigmoid function to dynamically suppress interference from reflective areas and enhance the saliency of defect edge features in the spatial dimension.

[0017] In a preferred embodiment, in step S4, SE-Net is embedded in the C2f module of the Neck network. The input feature map is compressed to a 1×1×C dimension through global average pooling, and after passing through two fully connected layers and the ReLU activation function, the channel weight coefficients are output. The weight coefficients are calibrated using the Sigmoid function and multiplied channel by channel with the original feature map. The channel compression ratio is 16:1, and the computational cost increases by only 3%, but the efficiency of transmitting key defect channel information is improved by 22%, effectively reducing the false detection rate.

[0018] In a preferred embodiment, in step S5, BiFPN is used instead of the traditional PANet structure to construct a bidirectional cross-layer connection network. The input feature maps are fused sequentially from high resolution to low resolution, and each layer introduces learnable weight coefficients initialized to 0.5. During feature fusion, the direct cross-layer connections between 80×80 and 20×20 resolution feature maps are preserved. After weighted fusion, a multi-scale feature pyramid is generated, improving the small target detection accuracy by 15% while maintaining the inference speed at 45 FPS.

[0019] In a preferred embodiment, in step S6, the original CIoU loss function is replaced with the EIoU loss function, the center point distance penalty term is calculated using the squared Euclidean distance, and the width-to-height difference penalty term introduces the logarithmic function of the width-to-height ratio. The weight coefficient of the loss function is set to 0.05. During training, the gradient direction is dynamically adjusted to improve the convergence speed of the predicted bounding box center point by 30% and reduce the localization error of small cracks to within ±2 pixels.

[0020] In a preferred embodiment, in step S7, the Head network is designed with three independent detection blocks, corresponding to feature maps with resolutions of 80×80, 40×40, and 20×20, respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The bounding box branch uses a 3×3 convolutional kernel with 256 channels of output, and the classification branch uses a 1×1 convolutional kernel with 64 channels of output. For loss calculation, EIoU loss is used for bounding box regression, and Focal Loss is used for classification. The weight ratio of the two losses is 1:0.5, improving training efficiency by 20%.

[0021] In a preferred embodiment, in step S8, the input image is uniformly scaled to 640×640×3, and the normalization parameters are set to mean 0.485, 0.456, and 0.406, and standard deviation 0.229, 0.224, and 0.225. The model is initialized by loading pre-trained weights, with a training batch size of 16 and an initial learning rate of 0.001. A cosine annealing strategy is used to adjust the learning rate. After each training round, the validation set evaluation metrics include precision, recall, and mAP. The early stopping strategy is set to a patience value of 10 rounds. After model convergence, the test set mAP reaches 92.3%.

[0022] In a preferred embodiment, in step S9, the optimal model is loaded to perform inference on the test set, outputting the defect category, confidence level, and bounding box coordinates. The confidence threshold is set to 0.5, and the non-maximum suppression (IOU) threshold is 0.45. The detection results are overlaid onto the original borehole image using OpenCV. The bounding box color is differentiated according to the defect category, and the text annotation uses a 12-point font on a black background. The statistical false negative rate and false positive rate are less than 3.5% and 2.8%, respectively. The visualization results are exported in PNG format, supporting rapid re-inspection and quality assessment in industrial settings.

[0023] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, by replacing the backbone network with a bidirectional feature pyramid structure, the model can more efficiently fuse feature information at different scales, effectively capturing the deep features of micro-scale defects such as micron-level cracks and pores, and reducing the problem of missed detections caused by the small size or random distribution of defects. Simultaneously, the introduced attention mechanism, through dynamic adjustment of feature weights, suppresses the influence of interference factors such as blade surface reflection and oil stains on the detection results, making the identification of defect areas more accurate in complex backgrounds.

[0024] 2. In this invention, by optimizing the multi-scale feature fusion structure and simplifying redundant computation nodes, the model significantly reduces computational resource consumption while maintaining high accuracy, ensuring that the detection speed meets the real-time requirements of industrial scenarios. The decoupled detection head design allows classification and localization tasks to be processed independently, improving the model's parallel processing capability for multiple types of defects. At the same time, an improved loss function is used to optimize bounding box localization, reducing false detections and localization errors, and ensuring stable output of detection results.

[0025] 3. In this invention, by expanding sample diversity through data augmentation strategies, the model can better address the imaging noise issues caused by the high reflectivity of aerospace blade materials and their curved surface structures. After fusing the channel optimization module, the model's generalization ability for samples with different imaging quality defects is significantly improved. Even under low contrast or partial occlusion conditions, it can still maintain a high detection confidence level, providing reliable technical support for practical engineering applications. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the process principle of the present invention; Figure 2 This is a diagram showing the overall structure of the YOLOV8 model in this invention; Figure 3 This is a structural diagram of the CBAM_C2f module in this invention; Figure 4 This is a structural diagram of the SENet_C2f module in this invention; Figure 5 This is a structural diagram of the BiFPN module in this invention; Figure 6 This is a diagram of the convolution block structure in this invention; Figure 7 This is a diagram of the spatial pyramid pooling layer structure in this invention; Figure 8 This is a structural diagram of the detection block in this invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] Reference Figure 1-8 , A method for detecting defects in aero-engine blades based on an improved YOLOv8 model, the method comprising the following steps: S1: Collect borehole images of aero-engines as raw data, and use annotation tools to annotate the bounding boxes of defects such as cracks, corrosion, and porosity. Divide the data into training set, validation set and test set, and perform data augmentation by rotation, scaling, translation, inversion and contrast adjustment to expand the sample diversity.

[0029] S2: The original backbone network is replaced with a bidirectional feature pyramid network (BiFPN). By optimizing multi-scale feature fusion through cross-scale connections and learnable weights, the ability to extract deep features of small defects is enhanced, while reducing redundant nodes to reduce computational cost.

[0030] S3: Add a convolutional block attention module (CBAM) before the C2f module of the Backbone network. Dynamically adjust feature weights through channel attention and spatial attention branches to suppress surface reflection and oil pollution interference, and improve the saliency of defect spatial features.

[0031] S4: Embed SE-Net into the C2f module of the Neck network, compress the channel dimension of the feature map through global pooling, and recalibrate the channel weights using fully connected layers and activation functions to enhance the information transmission of key defect channels and improve the generalization of the model.

[0032] S5: BiFPN is used to replace the traditional PANet structure. It fuses feature maps of different resolutions through bidirectional cross-layer connections and combines a weighted feature pyramid to enhance the detection effect of small target defects and reduce the consumption of computing resources.

[0033] S6: Replace the original CIoU loss function with the EIoU loss function. By minimizing the distance between the center point of the predicted box and the true box and penalizing the difference in width and height, the positioning accuracy of irregular defects such as cracks and pores is improved.

[0034] S7: Design three independent detection blocks in the Head network, corresponding to 80×80, 40×40 and 20×20 resolutions respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The loss is output through independent convolutional layers to improve detection efficiency.

[0035] S8: Scale the preprocessed image to 640×640×3 and normalize it. Load the improved YOLOV8 model for parameter initialization. Use an iterative training strategy. After each training round, use the validation set to evaluate the precision and recall until the model converges.

[0036] S9: Load the optimal parameter model to perform inference on the test set, output the defect category, confidence level and bounding box coordinates, combine the borehole images to generate visualized detection results, and calculate the false negative rate and false positive rate to verify the practicality of the model in complex industrial scenarios.

[0037] In step S1, borehole images of aero-engines are collected as raw data, covering typical defect types such as cracks, corrosion, and porosity. Annotation tools are used to accurately label the defect regions in each image, ensuring a consistent annotation format. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Data augmentation employs rotation angles of ±30 degrees, scaling ratios of 0.8 to 1.2 times, horizontal and vertical translation ranges of ±15%, mirror reversal, and contrast adjustment ranges of 0.5 to 1.5 times. Diverse samples are generated through random combinations of augmentation strategies to improve the model's robustness to noise and complex backgrounds.

[0038] In step S2, the original backbone network is replaced with a bidirectional feature pyramid network (BiFPN). Feature maps of different resolutions are fused through four layers of cross-scale connections, with learnable weight coefficients introduced in each layer to dynamically adjust feature importance. The input image size is fixed at 640×640×3, and the BiFPN output feature map sizes are 80×80, 40×40, and 20×20, respectively. Removing redundant nodes reduces computation by 18%, while retaining cross-layer skip connections to enhance the efficiency of transferring small defect features.

[0039] In step S3, a convolutional block attention module (CBAM) is embedded before the C2f module of the backbone network. The channel attention branch generates channel weights through global average pooling and fully connected layers, while the spatial attention branch generates a spatial weight matrix using 7×7 convolutional kernels. After the feature map is weighted by the two branches, the channel dimension is compressed to 1 / 16 of its original size. The spatial weights are normalized using the sigmoid function to dynamically suppress interference from reflective areas and improve the saliency of defect edge features in the spatial dimension.

[0040] In step S4, SE-Net is embedded into the C2f module of the Neck network. The input feature map is compressed to a 1×1×C dimension through global average pooling, and after passing through two fully connected layers and the ReLU activation function, the output channel weight coefficients are obtained. The weight coefficients are calibrated using the Sigmoid function and multiplied channel by channel with the original feature map. The channel compression ratio is 16:1, and the computational cost increases by only 3%, but the efficiency of information transmission in key defect channels is improved by 22%, effectively reducing the false detection rate.

[0041] In step S5, BiFPN is used to replace the traditional PANet structure to construct a bidirectional cross-layer connection network. The input feature maps are fused sequentially from high resolution to low resolution, and each layer introduces learnable weight coefficients initialized to 0.5. During feature fusion, the direct cross-layer connection between 80×80 and 20×20 resolution feature maps is preserved. After weighted fusion, a multi-scale feature pyramid is generated, improving the accuracy of small target detection by 15% while maintaining the inference speed at 45 FPS.

[0042] In step S6, the original CIoU loss function is replaced with the EIoU loss function. The center point distance penalty term is calculated using the squared Euclidean distance, and the width-to-height difference penalty term introduces the logarithm of the width-to-height ratio. The weight coefficient of the loss function is set to 0.05. During training, the gradient direction is dynamically adjusted to improve the convergence speed of the predicted bounding box center point by 30% and reduce the localization error of small cracks to within ±2 pixels.

[0043] In step S7, the Head network is designed with three independent detection blocks, corresponding to feature maps with resolutions of 80×80, 40×40, and 20×20, respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The bounding box branch uses a 3×3 convolutional kernel with 256 channels of output, while the classification branch uses a 1×1 convolutional kernel with 64 channels of output. For loss calculation, EIoU loss is used for bounding box regression, and Focal Loss is used for classification. The weight ratio of the two losses is 1:0.5, improving training efficiency by 20%.

[0044] In step S8, the input images are uniformly scaled to 640×640×3, and the normalization parameters are set to mean 0.485, 0.456, and 0.406, and standard deviation 0.229, 0.224, and 0.225. The model is initialized by loading pre-trained weights, with a training batch size of 16 and an initial learning rate of 0.001. Cosine annealing is used to adjust the learning rate. After each training round, the validation set evaluation metrics include precision, recall, and mAP. The early stopping strategy is set to a patience value of 10 rounds. After model convergence, the test set mAP reaches 92.3%.

[0045] In step S9, the optimal model is loaded to perform inference on the test set, outputting the defect category, confidence score, and bounding box coordinates. The confidence score threshold is set to 0.5, and the non-maximum suppression (IOU) threshold is 0.45. The detection results are overlaid onto the original borehole image using OpenCV. The bounding box color is differentiated according to the defect category, and the text annotation uses a 12-point font on a black background. The statistical false negative rate and false positive rate are less than 3.5% and 2.8%, respectively. The visualization results are exported in PNG format, supporting rapid re-inspection and quality assessment in industrial settings.

[0046] From the above, we can conclude that: In this invention, by replacing the backbone network with a bidirectional feature pyramid structure, the model can more efficiently fuse feature information at different scales, effectively capturing the deep features of micro-scale defects such as micron-level cracks and pores, and reducing the problem of missed detections caused by the small size or random distribution of defects. Simultaneously, the introduced attention mechanism, through dynamic adjustment of feature weights, suppresses the influence of interference factors such as blade surface reflection and oil stains on the detection results, making the identification of defect areas more accurate in complex backgrounds.

[0047] In this invention, by optimizing the multi-scale feature fusion structure and simplifying redundant computation nodes, the model significantly reduces computational resource consumption while maintaining high accuracy, ensuring that the detection speed meets the real-time requirements of industrial scenarios. The decoupled detection head design allows classification and localization tasks to be processed independently, improving the model's parallel processing capability for multiple types of defects. At the same time, an improved loss function is used to optimize bounding box localization, reducing false detections and localization errors, and ensuring stable output of detection results.

[0048] In this invention, by expanding sample diversity through data augmentation strategies, the model can better address the imaging noise issues caused by the high reflectivity of aerospace blade materials and their curved surface structures. After fusing the channel optimization module, the model's generalization ability for samples with different imaging quality defects is significantly improved. Even under low contrast or partial occlusion conditions, it can still maintain a high detection confidence level, providing reliable technical support for practical engineering applications.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0050] The foregoing description enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting defects in aero-engine blades based on an improved YOLOv8 model, characterized in that: The method includes the following steps: S1: Collect borehole images of aero-engines as raw data, use annotation tools to annotate the bounding boxes of cracks, corrosion and porosity defects, divide them into training set, validation set and test set, and perform data augmentation through rotation, scaling, translation, reversal and contrast adjustment to expand sample diversity; S2: The original backbone network is replaced with a bidirectional feature pyramid network BiFPN. Multi-scale feature fusion is optimized through cross-scale connections and learnable weights, which enhances the ability to extract deep features of small defects, while reducing redundant nodes to reduce computation. S3: Add a convolutional block attention module (CBAM) before the C2f module of the Backbone network. Dynamically adjust feature weights through channel attention and spatial attention branches to suppress surface reflection and oil interference, and improve the saliency of defect spatial features. S4: Embed SE-Net into the C2f module of the Neck network, compress the channel dimension of the feature map through global pooling, and recalibrate the channel weights using fully connected layers and activation functions to strengthen the information transmission of key defect channels and improve the generalization of the model. S5: BiFPN is adopted to replace the traditional PANet structure. It fuses feature maps of different resolutions through bidirectional cross-layer connections and combines weighted feature pyramids to enhance the detection effect of small target defects and reduce the consumption of computing resources. S6: Replace the original CIoU loss function with the EIoU loss function. By minimizing the distance between the center point of the predicted box and the true box and penalizing the difference in width and height, the localization accuracy of cracks and irregular defects with pores is improved. S7: Design three independent detection blocks in the Head network, corresponding to 80×80, 40×40 and 20×20 resolutions respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The loss is output through independent convolutional layers to improve detection efficiency. S8: Scale the preprocessed image to 640×640×3 and normalize it. Load the improved YOLOV8 model to initialize the parameters. Use an iterative training strategy. After each training round, use the validation set to evaluate the precision and recall until the model converges. S9: Load the optimal parameter model to perform inference on the test set, output the defect category, confidence level and bounding box coordinates, combine the borehole images to generate visualized detection results, and calculate the false negative rate and false positive rate to verify the practicality of the model in complex industrial scenarios.

2. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S1, borehole images of aero-engines are collected as raw data, covering typical defect types such as cracks, corrosion, and porosity. An annotation tool is used to accurately annotate the defect areas in each image to ensure that the annotation format is uniform. The dataset is divided into training set, validation set, and test set in a ratio of 7:2:

1.

3. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S2, the original backbone network is replaced with a bidirectional feature pyramid network (BiFPN). Feature maps of different resolutions are fused through four layers of cross-scale connections, and learnable weight coefficients are introduced in each layer to dynamically adjust the importance of features.

4. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S3, a convolutional block attention module (CBAM) is embedded before the C2f module of the Backbone network. The channel attention branch generates channel weights through global average pooling and fully connected layers, and the spatial attention branch generates a spatial weight matrix using a 7×7 convolutional kernel.

5. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S4, SE-Net is embedded in the C2f module of the Neck network. The input feature map is compressed to 1×1×C dimension through global average pooling, and after passing through two fully connected layers and the ReLU activation function, the channel weight coefficients are output. The weighting coefficients are calibrated using the Sigmoid function and multiplied channel by channel with the original feature map, with a channel compression ratio of 16:

1.

6. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S5, BiFPN is used to replace the traditional PANet structure to construct a bidirectional cross-layer connection network. The input feature maps are fused sequentially from high resolution to low resolution, and each layer introduces learnable weight coefficients initialized to 0.

5. During feature fusion, the direct cross-layer connections between 80×80 and 20×20 resolution feature maps are preserved, and a multi-scale feature pyramid is generated after weighted fusion.

7. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S6, the original CIoU loss function is replaced with the EIoU loss function, the center point distance penalty term is calculated using the square of the Euclidean distance, and the width-height difference penalty term introduces the width-height ratio logarithm function.

8. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S7, the Head network is designed with three independent detection blocks, corresponding to feature maps with resolutions of 80×80, 40×40, and 20×20, respectively. Each detection head is decoupled into a bounding box prediction branch and a classification prediction branch. The bounding box branch uses a 3×3 convolutional kernel and 256-channel output, while the classification branch uses a 1×1 convolutional kernel and 64-channel output.

9. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S8, the input image is uniformly scaled to 640×640×3, and the normalization parameters are set to mean 0.485, 0.456, 0.406, and standard deviation 0.229, 0.224, 0.

225. The model is initialized by loading pre-trained weights, the training batch size is set to 16, the initial learning rate is 0.001, and the learning rate is adjusted using a cosine annealing strategy.

10. The method for detecting defects in aero-engine blades based on the improved YOLOv8 model as described in claim 1, characterized in that: In step S9, the optimal model is loaded to perform inference on the test set, and the defect category, confidence level and bounding box coordinates are output. The confidence level threshold is set to 0.5 and the non-maximum suppression IOU threshold is 0.

45. The detection results are superimposed onto the original borehole image through OpenCV. The bounding box color is distinguished according to the defect category, and the text annotation uses a 12-point font and a black background.

Images