Aircraft skin defect multi-scale detection system and method based on improved YOLOv11

Abstract

The invention relates to the technical field of aircraft skin defect detection, and discloses an improved YOLOv11-based aircraft skin defect multi-scale detection system and method. Extracting multi-scale features of the input image through HMSB; the HMSB is composed of a plurality of parallel sub-networks, each sub-network independently processes input images with different resolutions, and a lightweight Transform encoder is adopted to extract multi-scale features; the multi-scale features are adaptively fused through DFFM; the DFFM calculates the attention weight of each scale feature through a gating mechanism to realize self-adaptive weighted fusion of the multi-scale features; the defects are classified and positioned through SDH, and a detection result is output; the SDH comprises detection branches customized for various defect types, and each branch adopts a deformable convolution and a Transform decoder to improve the positioning precision of irregular defects. According to the invention, through hierarchical multi-scale feature processing and a dynamic adaptive fusion mechanism, the performance bottleneck of a traditional method in detection of tiny and irregular defects is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft skin defect detection, and specifically to an aircraft skin defect multi-scale detection system and method based on improved YOLOv11. Background Technique

[0002] Aircraft skin defect detection is a key link in aviation safety maintenance. Traditional methods relying on manual visual inspection have problems such as low efficiency, strong subjectivity, and difficulty in covering hidden areas. In recent years, detection technologies based on deep learning have significantly improved the automation level. Among them, the YOLO (You Only Look Once) series of algorithms have been widely used in industrial defect detection due to their real-time advantages. However, in resource-constrained scenarios such as remote airports, the existing methods still face the following challenges: 1) Missed detection of small target defects: Conventional single-scale detection networks are not sensitive enough to tiny cracks or corrosion. Although multi-scale methods such as FPN (Feature Pyramid Network) can alleviate this problem, their fixed-weight fusion mechanism is difficult to adapt to the morphological diversity of aircraft skin defects.

[0003] 2) Computational resource limitations: Remote airports often lack high-performance computing devices. Existing lightweight networks such as Mobile-YOLO reduce the computational amount, but sacrifice the fine-grained fusion ability of multi-scale features.

[0004] 3) Poor adaptability to dynamic environments: Changes in lighting and complex background interference lead to a high false detection rate of traditional methods. Dynamic feature fusion technologies such as ASFF (Adaptive Spatial Feature Fusion) have not been deeply integrated with the multi-scale detection framework.

[0005] In the current improvement schemes, the open-source project Hi Fuse (Hierarchical Multi-Scale Feature Fusion Network for Medical Image Classification) fuses multi-scale features through a parallel hierarchical structure, but it is not optimized for defect detection; although YOLO-FDD (an efficient aircraft skin fastener defect detection network) introduces an attention mechanism to enhance the positioning accuracy, its single-path feature extraction limits the small target detection performance. In addition, the modular design in SLGA-YOLO (fused enhanced attention mechanism and efficient self-architecture lightweight YOLO) is only used for model compression and does not give full play to the collaborative advantages of sub-networks.

[0006] The limitations of the existing technology can be summarized as: 1) Static fusion strategy: Most multi-scale networks (such as FPN, Fuzzy Petri Nets) adopt predefined fusion rules and cannot dynamically adjust feature contributions according to the input image.

[0007] 2) Computational redundancy: Hierarchical structures such as MNN-CH (Modular neural network via exploring category hierarchy) improve the degree of modularity but do not optimize the computational load for edge devices.

[0008] 3) Insufficient generalization ability: Methods that improve YOLOv8n and others rely on specific datasets and are difficult to adapt to the complex environments of remote airports. Summary of the Invention

[0009] Aiming at the above problems, the purpose of the present invention is to provide an aircraft skin defect multi-scale detection system and method based on improved YOLOv11. Through hierarchical multi-scale feature processing and dynamic adaptive fusion mechanism, the performance bottleneck of traditional methods in detecting tiny and irregular defects is solved. The technical solutions are as follows: An aircraft skin defect multi-scale detection system based on improved YOLOv11 includes: Image acquisition system: Obtain the original aircraft skin image through a drone or a fixed camera; Data preprocessing unit: Perform enhancement and noise reduction processing on the original aircraft skin image; HMS-YOLOv11 module: Used to achieve multi-scale feature extraction and dynamic fusion; including a hierarchical multi-scale backbone network, a dynamic feature fusion module, and a dedicated detection head; The hierarchical multi-scale backbone network is composed of multiple parallel sub-networks. Each sub-network independently processes input images of different resolutions and uses a lightweight Transformer encoder to extract multi-scale features; The dynamic feature fusion module calculates the attention weights of features at each scale through a gating mechanism to achieve adaptive weighted fusion of multi-scale features; The dedicated detection head contains detection branches customized for various defect types. Each branch uses deformable convolution and a Transformer decoder to improve the localization accuracy of irregular defects.

[0010] An aircraft skin defect multi-scale detection method based on improved YOLOv11 includes the following steps: Step 1: Obtain the original aircraft skin image through a drone or a fixed camera and perform enhancement and noise reduction processing; Step 2: Extract multi-scale features of the input image through a hierarchical multi-scale backbone network; the hierarchical multi-scale backbone network is composed of multiple parallel sub-networks, each sub-network independently processes input images of different resolutions, and a lightweight Transformer encoder is used to extract multi-scale features; Step 3: Adaptive fusion of multi-scale features through a dynamic feature fusion module; the dynamic feature fusion module calculates the attention weights of each scale feature through a gating mechanism to achieve adaptive weighted fusion of multi-scale features; Step 4: Classify and locate defects through a dedicated detection head and output the detection results; the dedicated detection head contains detection branches customized for various defect types, and each branch uses deformable convolution and a Transformer decoder to improve the localization accuracy of irregular defects; Step 5: Jointly optimize the classification loss , regression loss and attention consistency loss , and complete end-to-end training.

[0011] The beneficial effects of the present invention are as follows: 1) The present invention adopts multi-scale specialization processing: the detection capabilities of defects of different sizes are independently optimized through parallel sub-networks to achieve scale specialization processing, and it can independently process defects of different scales to avoid feature confusion (such as feature conflicts between micro-cracks and large-area corrosion); at the same time, the customized design of the dedicated detection head improves the detection accuracy of specific defects.

[0012] 2) The present invention adopts a dynamic multi-scale fusion mechanism: a learnable scale weight (DFFM) is introduced into aircraft defect detection for the first time to replace fixed fusion rules such as FPN; the weights of multi-scale features are dynamically adjusted, which is better than the fixed attention template of YOLO-FDD.

[0013] 3) The present invention adopts a resource-aware design: the modular architecture allows selective update of sub-networks on edge devices, significantly reducing the hardware maintenance cost of remote airports, and the lightweight Transformer encoder reduces the number of parameters by 70%; the modular update mechanism only needs to retrain a single sub-network (such as the crack detection module); it makes up for the deficiency of SLGA-YOLO that only compresses the model.

[0014] 4) The present invention achieves environmental robustness: the dynamic fusion mechanism can adapt to interferences such as changes in illumination and complex backgrounds, improving the detection stability in actual scenarios.

[0015] 5) The present invention adopts a unified training protocol to jointly optimize the classification loss , regression loss and attention consistency loss , improving the stability of the model. Brief Description of the Drawings

[0016] Figure 1 It is an architecture diagram of the integrated HMS-YOLOv11 module for the aircraft surface inspection system (ASIS).

[0017] Figure 2 It is a detailed internal structure diagram of the HMS-YOLOv11 module. Detailed Implementation Manner

[0018] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0019] The hierarchical multi-scale modular YOLOv11 (HMS-YOLOv11, Hierarchical multi-scale modularization-YOLOv11) proposed by the present invention is an innovative architecture designed specifically for aircraft skin defect detection. Its core value lies in solving the performance bottleneck of traditional methods in detecting tiny and irregular defects through hierarchical multi-scale feature processing and dynamic adaptive fusion mechanism.

[0020] As Figure 1 shown, the detection system of the present invention is composed of the aircraft surface inspection system (ASIS, Aircraft surface inspection system) integrated with the HMS-YOLOv11 module. The architecture diagram includes:

[0021] Image acquisition system (IAS, Image acquisition system): Responsible for obtaining the original aircraft skin image through an unmanned aerial vehicle or a fixed camera.

[0022] Camera or Drone (Camera or UAV): Image acquisition device, outputting the unprocessed original image.

[0023] High Resolution Sensor (High-resolution sensor): Ensuring the complete capture of image details.

[0024] Data preprocessing unit (DPU, Data preprocessing unit): Performing enhancement and noise reduction processing on the original image. Image Enhancer (Image enhancer): Enhancing the image contrast and brightness to optimize the subsequent detection effect.

[0025] Noise Reducer (Noise reducer): Eliminating image noise and reducing environmental interference.

[0026] HMS-YOLOv11 module: The core detection module, realizing multi-scale feature extraction and dynamic fusion.

[0027] Hierarchical Multi Scale Backbone (HMSB): Processes images of different scales in parallel and extracts multi-resolution features.

[0028] Dynamic Feature Fusion Module (DFFM): Adaptively fuses multi-scale features to enhance the representation of key regions.

[0029] Specialized Detection Heads (SDH): Specialized detection heads for defects such as cracks, dents, and corrosion.

[0030] Result visualization and reporting unit (RVRU): Generates visualization results and maintenance reports.

[0031] Defect Visualizer: Labels the location and type of defects and outputs a marked image. Report Generator: Generates a structured report for maintenance personnel to use.

[0032] User Interface: Displays the detection results and reports.

[0033] The HMS-YOLOv11 module replaces the traditional single-scale detection network and significantly improves the detection accuracy and efficiency through dynamic feature fusion and specialized detection heads. The internal structure of the HMS-YOLOv11 module is as Figure 2 shown and consists of three core components:

[0034] Hierarchical Multi Scale Backbone (HMSB): After receiving the input image, it processes the high, medium, and low-resolution versions through three parallel sub-networks respectively. Each sub-network uses a lightweight Transformer encoder to extract scale-specific features.

[0035] Dynamic Feature Fusion Module (DFFM): Calculates the attention weights of features at each scale through a gating mechanism to achieve adaptive weighted fusion of feature maps.

[0036] Specialized Detection Heads (SDH): Consists of three branches, designed respectively for cracks (Crack Head), dents (DentHead), and corrosion (Corrosion Head), and uses deformable convolution and a Transformer decoder to improve the localization accuracy of irregular defects.

[0037] The system workflow is as follows: Input Image → HMSB Multi-scale Feature Extraction → DFFM Dynamic Fusion → SDH Defect Classification and Localization. All components are jointly optimized through a unified training protocol, supporting end-to-end deployment.

[0038] Figure 2 Among them: Input Image (Input Image): Enters the multi-scale processing workflow.

[0039] Hierarchical Multi-scale Backbone Network (HMSB): Processes image inputs at different scales in parallel.

[0040] Scale 1 Sub Network: Processes the original resolution image and captures micro-defect features.

[0041] Scale 2 Sub Network: Processes medium-resolution images and balances details and computational efficiency.

[0042] Scale K Sub Network: Processes low-resolution images and focuses on large-scale defects.

[0043] Dynamic Feature Fusion Module (DFFM): Adaptively fuses multi-scale features.

[0044] Attention Weight Calculator: Calculates the weights of features at each scale and dynamically adjusts the fusion ratio.

[0045] Feature Fusion Operator: Performs weighted fusion and outputs an optimized feature map.

[0046] Specialized Detection Head (SDH): Detectors for different defect types.

[0047] Crack Detector: Optimizes the localization of micro linear defects.

[0048] Dent Detector: Adapts to the recognition of irregular concave areas.

[0049] Corrosion Detector: A corrosion detector that processes defects with large-area texture changes.

[0050] Unified Training Protocol: Jointly optimizes the parameters of all modules. Receives the loss feedback from each detector and updates the parameters of the hierarchical multi-scale backbone network, dynamic feature fusion module, and specialized detection head.

[0051] The dynamic feature fusion module realizes the adaptive fusion of multi-scale features through the attention mechanism, overcoming the limitations of traditional fixed-weight fusion; the dedicated detection head is customized for different defect types, significantly improving the detection accuracy.

[0052] 1. Hierarchical multi-scale backbone network (HMSB); For the input image , HMSB generates three groups of scale feature maps: ; Among them, (original image), , correspond to the high, medium, and low-resolution paths respectively. Each path extracts features through a lightweight Transformer encoder : ; Among them, are the sub-network parameters, is the unified number of channels. The encoder uses depthwise separable convolutions to reduce the computational load, and its multi-head attention mechanism calculation is: ; Among them, are the query, key, and value matrices, is the dimensional scaling factor.

[0053] 2. Dynamic feature fusion module (DFFM); DFFM calculates the weights of each scale through a gating unit: ; Among them, is the global average pooling, and are learnable parameters. The fused feature is obtained through weighted upsampling: ; Among them, is the transposed convolution operation with a stride to ensure that all feature maps are restored to the original image resolution.

[0054] 3. Dedicated detection head (SDH); Each detection head contains: (1) Deformable convolutional layer: adapting to irregular defect shapes through offsets : ; Among them, is the receptive field of the convolutional kernel, which is predicted by the auxiliary network.

[0055] (2) Transformer decoder: Interacts the fused features with the learnable query vector to output the defect location and confidence .

[0056] 4. Loss function; Jointly optimize the classification loss , regression loss and attention consistency loss to complete end-to-end training. 5. Specific embodiments (1) Lightweight deployment solution based on edge computing: This embodiment describes the optimized deployment method of HMS-YOLOv11 on the NVIDIA Jetson AGX Orin edge device: 1) Selective loading of sub-networks: Dynamically load specific sub-networks in HMSB according to real-time detection requirements. For example, when the drone is flying at a low altitude, only the high-resolution path (Scale 1 Sub Network) is enabled, reducing the computational load by 60%.

[0058] 2) Quantization-aware training: Use INT8 quantization technology to compress the deformable convolutional layer in SDH, increasing the inference speed of the corrosion detector by 3 times and reducing the memory footprint to 40% of the original model.

[0059] 3) Asynchronous pipeline design: Decouple the feature fusion operation of DFFM from the detection task of SDH, and achieve parallel processing of multi-scale feature extraction and defect recognition through a double-buffer mechanism.

[0060] (2) Multi-modal data fusion detection scheme: This embodiment extends HMS-YOLOv11 to support multi-modal input of infrared and visible light images: 1) Cross-modal feature alignment: Add a cross-modal attention layer in HMSB to automatically align the feature maps of the visible light path (RGB SubNetwork) and the infrared path (Thermal Sub Network) in the channel dimension.

[0061] 2) Adaptive modal weight: DFFM is extended to a bi-modal version (Bi-DFFM), and the fusion weight of the infrared feature is dynamically adjusted through additional temperature sensor data, giving priority to using infrared features at night or in low-light conditions.

[0062] 3) Heterogeneous Detection Head Design: A new hot spot detector is added to the SDH to specifically process the abnormal temperature rise areas in infrared images, and it is used for complementary verification with the visible light detection results.

[0063] (3) Reconfigurable Modular Update Scheme: This embodiment provides two update strategies for the maintenance conditions of different airports: 1) Incremental Update: When a new corrosion type (such as stress corrosion cracking) is added, only the corrosion detector and its corresponding HMSB low-resolution path are retrained, and the parameters of other sub-networks are kept frozen. The amount of updated data is reduced to 15% of the full model.

[0064] 2) Federated Learning Collaborative Update: Edge devices at multiple remote airports collaboratively optimize the crack detector through encrypted gradient exchange, use local data to improve the generalization ability of the model, and avoid the centralized transmission of original data.

[0065] (4) Mobile Platform Dynamic Resolution Adaptation Scheme: This embodiment is applicable to the scenario of mobile shooting by drones: 1) Flight Altitude Adaptive Scaling: Dynamically adjust the input resolution scale factor of the HMSB according to the drone's GPS altitude data, s k . When the altitude exceeds 50 meters, automatically turn off the high-resolution path with s1 = 1.0 to save power.

[0066] 2) Region of Interest (ROI) Focus: Pre-screen the suspected defect areas through the on-board processor, and only input the ROI areas into the SDH for fine detection, which improves the overall processing frame rate by 2.2 times.

[0067] 3) Bandwidth Optimization Transmission: The fused feature map output by the DFFM is compressed by JPEG-XS and then transmitted back to the ground station, reducing the wireless transmission data volume by 80% while maintaining 95% detection accuracy.

[0068] (5) Extreme Environment Robustness Enhancement Scheme:

[0069] This embodiment improves the detection stability for harsh weather such as sand and dust, rain and snow: 1) Degradation-Aware Feature Compensation: Add a weather classifier at the front end of each sub-network of the HMSB. When sand and dust interference is recognized, automatically enhance the local contrast enhancement module of the middle layer path (Scale 2 Sub Network).

[0070] 2) Dynamic noise suppression: A noise mask based on meteorological sensor data is introduced in the DFFM fusion stage to suppress the interference of high-frequency noise caused by rain and snow on the calculation of attention weights.

[0071] 3) Multi-period model switching: HMS-YOLOv11 variants under different lighting conditions (such as morning / noon / night models) are pre-trained, and the model is seamlessly switched by the light intensity sensor.

[0072] In summary, the HMS-YOLOv11 system proposed by the present invention realizes a significant improvement in the accuracy, efficiency, and adaptability of aircraft skin defect detection through the collaborative design of a hierarchical multi-scale backbone network, a dynamic feature fusion module, and a dedicated detection head. The system innovatively solves the technical bottlenecks of traditional methods in small target detection, computing resource limitations, and environmental adaptability. Its core advantages are reflected in: 1) Multi-scale specialization processing ability: The detection performance of defects of different sizes is independently optimized through parallel sub-networks, avoiding the problem of feature confusion and significantly improving the synchronous detection accuracy of micro-cracks and large-scale corrosion.

[0073] 2) Dynamic adaptive mechanism: The DFFM module introduces learnable attention weights to achieve intelligent fusion of multi-scale features, overcoming the limitations of fixed fusion rules in complex scenarios.

[0074] 3) Edge-computing friendly design: The modular architecture supports selective loading and incremental update of sub-networks. Combined with quantization technology and asynchronous pipeline design, the system can operate efficiently on resource-constrained edge devices.

[0075] 4) Enhanced environmental robustness: Through multi-modal data fusion, dynamic resolution adaptation, and extreme environment optimization solutions, the system can adapt to diverse detection scenarios and harsh weather conditions at remote airports.

[0076] The technical solution of the present invention has been verified in actual airport maintenance scenarios. Compared with existing mainstream detection systems, while maintaining real-time performance (≥30FPS), the recall rate of micro-defects is increased to 92.3%, and the false alarm rate is reduced to 1.2%. The modular design reduces the system maintenance cost by 60%, which is particularly suitable for remote airports lacking professional maintenance teams. In the future, the multi-airport collaborative optimization ability can be further expanded through the federated learning framework to continuously improve the generalization performance of the model.

Claims

1. An aircraft skin defect multi-scale detection system based on improved YOLOv11, characterized in that, Including: Image acquisition system: Obtain the original aircraft skin images through drones or fixed cameras; Data preprocessing unit: Enhance and denoise the original aircraft skin images; HMS-YOLOv11 module: Used to achieve multi-scale feature extraction and dynamic fusion; Including a hierarchical multi-scale backbone network, a dynamic feature fusion module, and a dedicated detection head; The hierarchical multi-scale backbone network is composed of multiple parallel sub-networks. Each sub-network independently processes input images of different resolutions and uses a lightweight Transformer encoder to extract multi-scale features; The dynamic feature fusion module calculates the attention weights of features at each scale through a gating mechanism to achieve adaptive weighted fusion of multi-scale features; The dedicated detection head contains detection branches customized for various defect types. Each branch uses deformable convolution and a Transformer decoder to improve the localization accuracy of irregular defects.

2. The multi-scale detection system for aircraft skin defects based on the improved YOLOv11 according to claim 1, characterized in that, The hierarchical multi-scale backbone network includes three parallel sub-networks: High-resolution path: processing the original resolution image and extracting features through a lightweight Transformer encoder where H is the height of the original input image, , W is the width of the original input image, and C is the unified number of channels; represents a feature map with height H, width W, and 3 channels; ​ Medium-resolution path: Process the input image I2 with a resolution scale factor s2 = 0.5, and extract features through a lightweight Transformer encoder Extract features ; Low-resolution path: Process the input image I3 with a resolution scale factor s3 = 0.25, and extract features through a lightweight Transformer encoder Extract features ; The lightweight Transformer encoder uses depthwise separable convolution to reduce the computational amount, and its multi-head attention mechanism calculation is: ； Among them, are the query matrix, the key matrix, and the value matrix respectively, is the dimension scaling factor, T is the transpose symbol; Attention represents the calculation of the multi-head attention mechanism, and Softmax is the activation function.

3. The aircraft skin defect multi-scale detection system based on the improved YOLOv11 according to claim 2, characterized in that, The dynamic feature fusion module calculates the weights of each scale through a gating unit : ； Among them, is global average pooling, and are learnable parameters; is the feature extracted by the three-way parallel sub-network in the hierarchical multi-scale backbone network, ; represents the calculation of the gating unit; Fusion feature Obtained by weighted upsampling: ； Among them, is the step size of the deconvolution operation.

4. The aircraft skin defect multi-scale detection system based on the improved YOLOv11 according to claim 3, characterized in that, Each detection head of the dedicated detection head includes a deformable convolutional layer and a Transformer decoder; The deformable convolutional layer adapts to irregular defect shapes through offsets : ； Among them, is the receptive field of the convolutional kernel, predicted by the auxiliary network; is the weight of the convolutional kernel, is the input feature, is the coordinate of the pixel position currently being calculated on the output feature map, is the receptive field of the convolutional kernel is the relative position coordinate within; is the output feature of the deformable convolutional layer; The Transformer decoder will fuse the features with the learnable query vector to interact and output the defect location and confidence .

5. An aircraft skin defect multi-scale detection method based on improved YOLOv11, characterized in that Including the following steps: Step 1: Obtain the original aircraft skin images through drones or fixed cameras and perform enhancement and denoising processing; Step 2: Extract multi-scale features of the input images through the hierarchical multi-scale backbone network; The hierarchical multi-scale backbone network is composed of multiple parallel sub-networks. Each sub-network independently processes input images of different resolutions and uses a lightweight Transformer encoder to extract multi-scale features; Step 3: Adaptively fuse multi-scale features through the dynamic feature fusion module; The dynamic feature fusion module calculates the attention weights of features at each scale through a gating mechanism to achieve adaptive weighted fusion of multi-scale features; Step 4: Classify and locate defects through the dedicated detection head and output the detection results; The dedicated detection head contains detection branches customized for various defect types. Each branch uses deformable convolution and a Transformer decoder to improve the localization accuracy of irregular defects; Step 5: Jointly optimize the classification loss , regression loss and attention consistency loss to complete end-to-end training.

6. The multi-scale detection method for aircraft skin defects based on the improved YOLOv11 according to claim 5, characterized in that In step 2, for the original resolution image , the hierarchical multi-scale backbone network generates three groups of scale feature maps : ； Among them, is the resolution scale factor, corresponding to the high, medium, and low resolution paths respectively; Resize means processing through a hierarchical multi-scale backbone network; is the height of the original input image, is the width of the original input image; represents a feature map with height H, width W, and 3 channels; Each path passes through a lightweight Transformer encoder Extract features : ； Among them, is the sub-network parameter, is the unified number of channels; The encoder uses depthwise separable convolution to reduce the computational amount, and its multi-head attention mechanism calculation is: ； Among them, are the query matrix, the key matrix, and the value matrix respectively, is the dimension scaling factor; Attention represents the calculation of the multi-head attention mechanism, Softmax is the activation function; T is the transpose symbol.

7. The multi-scale detection method for aircraft skin defects based on the improved YOLOv11 according to claim 6, characterized in that, In step 3, the dynamic feature fusion module calculates the weights of each scale through a gating unit : ； Among them, is global average pooling, and are learnable parameters; is the feature extracted by the three-way parallel sub-network in the hierarchical multi-scale backbone network, ; Fusion feature Obtained by weighted upsampling: ； Among them, is the step size of the deconvolution operation.

8. The multi-scale detection method for aircraft skin defects based on the improved YOLOv11 according to claim 7, characterized in that Each detection head of the dedicated detection head in step 4 includes a deformable convolutional layer and a Transformer decoder; The deformable convolutional layer adapts to the irregular defect shape through the offset : ； Among them, is the receptive field of the convolutional kernel, predicted by the auxiliary network; is the weight of the convolutional kernel, is the input feature, is the coordinate of the pixel position currently being calculated on the output feature map, is the receptive field of the convolutional kernel The relative position coordinates within; the Transformer decoder will fuse the feature interact with the learnable query vector to output the defect position and confidence .

9. The method for multi-scale detection of aircraft skin defects based on the improved YOLOv11 according to claim 8, wherein In step 5, the loss function is defined as: ； where λ1, λ2, and λ3 are balance coefficients.

Images