An aircraft surface defect detection method

Abstract

The application discloses an aircraft surface defect detection method, comprising the following steps: according to the obtained aircraft skin surface image, carrying out nondestructive feature reorganization through a space-depth convolution module to obtain a first feature map; according to the first feature map, carrying out multi-scale feature extraction and dual-domain selection enhancement through a self-adaptive selective edge enhancement module to obtain a second feature map; and according to the second feature map, carrying out dynamic optimization on boundary box regression through a Wise-CIoU loss function integrated with a non-monotonic focusing mechanism to obtain an aircraft surface defect detection result. The application can effectively retain sub-pixel information of tiny defects, suppress complex background interference, and improve boundary box positioning accuracy.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and in particular relates to a method for detecting defects on the surface of aircraft. Background Technology

[0002] The structural integrity of aircraft skin is a core foundation for ensuring aircraft service reliability and flight safety. In complex and variable service environments, the fuselage surface is susceptible to aerodynamic stress, chemical corrosion, and foreign object damage, leading to various surface defects such as cracks, scratches, and dents. Currently, aviation maintenance systems primarily rely on manual visual inspection and traditional non-destructive testing (NDT) techniques (such as ultrasonic testing, eddy current testing, and shear speckle interferometry). However, manual visual inspection results are heavily influenced by the subjective experience of inspectors, are inefficient, and involve risks associated with high-altitude operations. Traditional NDT techniques are sensitive to environmental vibrations, coupling agent quality, and the roughness of the tested surface, making it difficult to achieve closed-loop inspection in large-scale, real-time, and automated hangar or UAV inspection scenarios. In recent years, deep learning-based visual inspection methods have been gradually introduced into this field, using convolutional neural networks to achieve end-to-end defect identification, providing a new technical path for automated aircraft skin inspection.

[0003] However, aircraft skin defect detection still faces three major technical bottlenecks in practical applications: First, most surface defects (such as microcracks and shallow scratches) have a very low pixel ratio in the image and are tiny targets. The downsampling operations of traditional networks (stretch convolution or pooling) can easily lead to the annihilation of their fine-grained spatial information, resulting in missed detections. Second, the aircraft skin surface has strong background interference such as high metal reflectivity, irregular textures (such as rivets and coating particles) and uneven lighting. The contrast between defects and background is extremely low, and existing models cannot effectively decouple target features from redundant textures, resulting in a high false detection rate. Third, the actual defect sample distribution is extremely uneven, especially the scarcity of high-quality labeled samples. Traditional regression loss functions (such as CIoU) are based on static geometric constraints and are sensitive to low-quality outliers, which can easily cause training oscillations and limit the sub-pixel localization accuracy of the bounding box. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for detecting surface defects on aircraft, comprising: Based on the acquired aircraft skin surface image, the first feature map is obtained by lossless feature reconstruction through a spatial-depth convolution module. Based on the first feature map, a second feature map is obtained by performing multi-scale feature extraction and dual-domain selection enhancement through an adaptive selective edge enhancement module. Based on the second feature map, the bounding box regression is dynamically optimized by the Wise-CIoU loss function with integrated non-monotonic focusing mechanism to obtain the aircraft surface defect detection results.

[0005] Preferably, the process of obtaining the first feature map includes: The aircraft skin surface image is subjected to spatial-depth transformation, and multiple sub-feature maps are generated through periodic sampling; Multiple sub-feature maps are concatenated along the channel dimension to obtain an intermediate feature map with reduced spatial resolution and expanded channel number. The first feature map is obtained by performing cross-channel feature interaction and channel number mapping on the intermediate feature map through a convolutional layer with a stride of 1.

[0006] Preferably, the process of obtaining the second feature map includes: The first feature map is input into the adaptive selective edge enhancement module, and multi-scale features and local detail features are extracted through parallel processing of multi-scale perception path and local perception path, respectively. The multi-scale features are subjected to edge enhancement processing to obtain the multi-scale features with enhanced edges; The local detail features and the edge-enhanced multi-scale features are concatenated along the channel dimension to obtain the fused features; The fused features are globally evaluated and filtered in the spatial and frequency domains using a dual-domain selection mechanism to obtain purified features. The purified features are subjected to channel dimensionality reduction and recombination to obtain the second feature map.

[0007] Preferably, the process of edge enhancement processing of the multi-scale features includes: Average pooling is performed on the input feature map to extract low-frequency background information; Subtract the low-frequency background information from the input feature map to extract high-frequency information containing edge details; The high-frequency information is passed through a convolutional layer with a Sigmoid activation function to generate an edge attention weight map; The edge attention weight map is superimposed back onto the input feature map to complete the adaptive activation and enhancement of edge features.

[0008] Preferably, the construction process of the multi-scale sensing path includes: By using multiple parallel adaptive average pooling layers, local semantic signals at different spatial scales are captured to obtain multi-scale pooling features. The multi-scale pooling features are convolved and then upsampled to restore the original resolution to obtain the multi-scale features.

[0009] Preferably, the Wise-CIoU loss function is obtained by multiplying the CIoU loss term by the non-monotonic focusing coefficient; The non-monotonic focusing coefficient is dynamically determined based on the outlier degree between the predicted box and the true box.

[0010] Preferably, the outlier is defined as the ratio of the IoU loss of the predicted box to the average IoU loss of all predicted boxes; When the outlier value represents a high-quality sample, the non-monotonic focusing coefficient is assigned a gradient weight that is higher than a preset threshold range. When the outlier value represents an outlier sample, the non-monotonic focusing coefficient decreases non-monotonically to weaken gradient interference.

[0011] Compared with the prior art, the present invention has the following advantages and technical effects: First, this invention constructs a lossless feature reconstruction mechanism based on spatial-depth convolution (SPD-Conv), replacing traditional stride convolution with spatial-depth transformation, which physically curbs the sub-pixel-level information annihilation of tiny defects during downsampling, and significantly improves the recall rate of extremely fine cracks and scratches. Secondly, this invention designs an adaptive selective edge enhancement (ASE) module that integrates a dual-domain selection mechanism. By using an edge enhancer to compensate for high-frequency signals and combining spatial and frequency domain collaborative screening, it effectively suppresses strong background noise and redundant texture interference on the skin surface, and significantly reduces the false detection rate. Finally, this invention introduces the Wise-CIoU loss function with an integrated non-monotonic focusing mechanism to dynamically adjust the gradient contribution rate of samples of different quality, thereby mitigating the interference of outliers on the training process and improving the sub-pixel accuracy and convergence stability of bounding box regression. Experimental results show that compared with the baseline YOLO11 model, the present invention improves mAP50 by 7.15% and mAP50-95 by 2.43%, achieving accurate boundary alignment and high-reliability detection, especially in the categories of fine cracks and scratches. Attached Figure Description

[0012] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a detailed architecture diagram of MsDEE-Net according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the SPD-Conv module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of ASE according to an embodiment of the present invention; Figure 4 This is a visualization of the YOLO11 baseline model in aircraft skin surface defect detection according to an embodiment of the present invention. Figure 5 This is a visualization of the MsDEE-Net application in aircraft skin surface defect detection according to an embodiment of the present invention. Detailed Implementation

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0015] like Figure 1 As shown, this embodiment provides a method for detecting surface defects on an aircraft, including: Based on the acquired aircraft skin surface image, the first feature map is obtained by lossless feature reconstruction through a spatial-depth convolution module. Based on the first feature map, a second feature map is obtained by performing multi-scale feature extraction and dual-domain selection enhancement through an adaptive selective edge enhancement module. Based on the second feature map, the bounding box regression is dynamically optimized by integrating the Wise-CIoU loss function with a non-monotonic focusing mechanism to obtain the aircraft surface defect detection results.

[0016] To address the multiple challenges of ASD (Advanced Surface Defect) detection, including fine-scale defects, extremely low contrast, and severe background texture interference, this embodiment proposes a multi-scale discriminative edge enhancement network (MsDEE-Net) based on an improved YOLO11. First, to address the issue of vanishing features in deep networks due to small defects, a lossless reconstruction mechanism based on spatial-depth convolution (SPD-Conv) is constructed to preserve sub-pixel topological information through spatial-depth transformation. Second, an adaptive selective edge-enhancement (ASE) module is designed, and a dual-domain selection mechanism (DSM) is innovatively integrated to more effectively suppress non-target redundant information on the skin surface. Finally, to enhance the model's localization stability at high IoU thresholds, a Wise-CIoU loss function integrating a non-monotonic focusing mechanism is constructed. Experimental results show that the solution in this embodiment performs excellently, with mAP50 improved by 7.15% compared to the baseline, and overall positioning accuracy (mAP50-95) improved by 2.43%. Especially in the detection of small cracks and scratches, it demonstrates more accurate boundary alignment capability and higher industrial inspection reliability.

[0017] Compared to existing solutions, the MsDEE-Net constructed in this embodiment establishes a complete end-to-end closed loop, encompassing feature preservation, information filtering, and detection output. This systematic design aims to address the issue of accuracy loss across the entire chain from signal acquisition to decision output in complex industrial environments.

[0018] Furthermore, the process of obtaining the first feature map includes: A space-depth transformation is performed on the aircraft skin surface image, and multiple sub-feature maps are generated through periodic sampling; Multiple sub-feature maps are concatenated along the channel dimension to obtain an intermediate feature map with reduced spatial resolution and expanded channel number. The first feature map is obtained by performing cross-channel feature interaction and channel number mapping on the intermediate feature map through a convolutional layer with a stride of 1.

[0019] Furthermore, this embodiment constructs a lossless feature reconstruction mechanism based on SPD-Conv. Addressing the challenge of information annihilation during downsampling due to minute defects, full-space fidelity downsampling based on SPD-Conv is performed. By replacing traditional stride convolution with spatial-depth transformation, the loss of sub-pixel-level signals is physically prevented, providing a highly complete topological foundation for subsequent feature extraction.

[0020] Specifically, to address the three core pain points in aircraft skin defect detection—loss of extremely small targets, background texture interference, and unstable positioning accuracy—this embodiment proposes a deep learning framework, MsDEE-Net, based on an improved YOLO11. The core design principles of this network are information fidelity and selective enhancement. Its overall architecture is as follows: Figure 1 As shown, MsDEE-Net employs an SPD-Conv structure at the bottom layer of its backbone network. Through spatial-to-channel domain conversion, it avoids the loss of micro-crack information during the initial downsampling stage. In the feature fusion network, this embodiment innovatively designs an ASE module. This module captures global and local features of the skin surface through parallel multi-scale paths, extracts high-frequency signals, and filters redundant textures. In the prediction head, MsDEE-Net integrates the Wise-CIoU loss function, which improves localization robustness in complex backgrounds by dynamically balancing the regression gradients of samples of different quality.

[0021] When detecting defects smaller than 1 mm, traditional strided convolution or pooling operations can lead to excessive compression or even loss of fine-grained spatial information. This study argues that later feature fusion alone cannot compensate for the irreversible information loss in the shallow layers of the early network. Therefore, MsDEE-Net introduces the SPD-Conv module.

[0022] This mechanism achieves lossless downsampling through space-depth transformation. Given an input tensor... (in For spatial resolution, (where the number of channels is 2) The SPD operation periodically samples the spatial dimension by a set step size (set to 2 in this embodiment) to generate 4 sub-feature maps. The mapping relationship can be expressed as: here, Indicates from index Initially, samples were taken with a step size of 2. Through the above slicing operation, the network extracts... , , , Four sub-feature maps, each with a dimension of 1. The module then concatenates these four sub-feature maps along the channel dimension: This operation downsamples the spatial dimension by a factor of 2, while expanding the channel dimension by a factor of 4, i.e., from... Convert to .

[0023] Finally, in order to achieve cross-channel feature interaction while preserving microscale damage signals, this embodiment introduces a step size of 1. Convolutional layers, such as Figure 2 As shown, the number of feature channels is reduced from Remapping to target number of channels : This non-destructive downsampling design ensures that discriminative signals from extremely small-scale damage can flow completely to the deep network and the detection head, laying a solid data foundation for subsequent accurate regression.

[0024] Furthermore, the process of obtaining the second feature map includes: The first feature map is input into the adaptive selective edge enhancement module, and multi-scale features and local detail features are extracted through parallel processing of multi-scale perception path and local perception path, respectively. Edge enhancement processing is performed on multi-scale features to obtain multi-scale features with enhanced edges; The local detail features and the multi-scale features enhanced by the edges are spliced ​​together in the channel dimension to obtain the fused features; A dual-domain selection mechanism is used to globally evaluate and filter the fused features in the spatial and frequency domains to obtain the purified features. The purified features are subjected to channel dimensionality reduction and recombination to obtain the second feature map.

[0025] Furthermore, the process of edge enhancement processing for multi-scale features includes: Average pooling is performed on the input feature map to extract low-frequency background information; Subtract the low-frequency background information from the input feature map to extract the high-frequency information containing edge details; High-frequency information is passed through a convolutional layer with a sigmoid activation function to generate an edge attention weight map; The edge attention weight map is superimposed back onto the input feature map to complete the adaptive activation and enhancement of edge features.

[0026] Furthermore, the construction process of the multi-scale sensing path includes: By using multiple parallel adaptive average pooling layers, local semantic signals at different spatial scales are captured to obtain multi-scale pooling features. Multi-scale pooling features are convolved separately and then upsampled to restore the original resolution to obtain multi-scale features.

[0027] Furthermore, this embodiment designs a dual-domain collaborative adaptive selective edge enhancement (ASE) module. It constructs a multi-scale perceptual path to actively capture local semantics at different physical sizes. Simultaneously, a dedicated edge enhancer is built-in, significantly improving representation capabilities. Finally, by fusing a dual-domain selection mechanism (DSM), deep coupling between spatial domain localization and frequency domain signal filtering is achieved, forcing the model to discriminatively decouple target features from redundant textures in strong background noise (such as metallic reflections).

[0028] Specifically, the ASE module is the core sensing unit of MsDEE-Net, and its detailed structure is as follows: Figure 3 As shown. This module aims to build a feature extraction system with multi-scale representation capabilities, and achieves high-sensitivity capture of minute damage to the skin surface by simulating the human visual perception logic of "first observing the outline, then distinguishing the details".

[0029] To address the receptive field mismatch issue caused by the variable shape of skin defects, the ASE module employs a dual-path parallel processing strategy. Given an input feature map... The module first utilizes A multi-scale sensing path is constructed using parallel adaptive average pooling layers: in Representing different spatial scales, the aim is to actively capture local semantic signals at different physical sizes. Subsequently, to compensate for the spatial information loss caused by pooling operations, the features at each scale are first upsampled to the original resolution using bilinear interpolation, denoted as . : Next, the module incorporates a dedicated edge booster (EdgeBoost, EB). For input features... First, use a The average pooling layer is used for smoothing to obtain low-frequency background information, and then high-frequency information containing edge details is extracted by residual subtraction. : Finally, the high-frequency features are fed into a convolutional layer with a sigmoid activation function to generate an edge attention weight map, which is then superimposed back onto the original features to complete the adaptive activation and enhancement of the edge features. in This represents the Sigmoid function. This difference operation can effectively extract high-frequency signals from the skin surface, such as fine scratches.

[0030] To remove artifact noise such as spectral reflectance from the skin surface, the ASE module introduces DSM after multi-scale feature stitching. Although DSM was originally designed for the highly challenging image restoration task of filtering out physical degradation noise, its inherent "signal and noise stripping" logic is highly compatible with the aviation inspection scenario.

[0031] In the DEE-Net architecture, this mechanism acts as a "feature purifier": its spatial domain branch is dedicated to suppressing uneven background interference, while its frequency domain branch uses waveform filtering to lock in high-frequency key information such as cracks and scratches. This paradigm shift from "pixel reconstruction" to "feature selection" significantly enhances the system's robustness in complex lighting environments.

[0032] Specifically, the network first concatenates the EdgeBoost-enhanced features Fboosti from each multi-scale branch with the features Flocal extracted from the local perceptual branch along the channel dimension to form the global fusion feature Fcat: Subsequently, DSM acts as a "feature purifier," performing global evaluation and filtering of Fcat simultaneously in both spatial and frequency dimensions. Finally, the purified features undergo channel dimensionality reduction and feature recombination via a 1×1 convolution, outputting the final result Fout from the ASE module: This dual-domain weighting mechanism can adaptively select key features that are highly relevant to the target task, and significantly optimize the expression accuracy of edge features while suppressing speckle noise.

[0033] Furthermore, the Wise-CIoU loss function is obtained by multiplying the CIoU loss term by the non-monotonic focusing coefficient; The non-monotonic focusing coefficient is dynamically determined based on the outlier degree between the predicted box and the true box.

[0034] Furthermore, this embodiment proposes a quality-aware regression optimization strategy. By constructing a Wise-CIoU loss function, it retains the comprehensive penalty term of CIoU on overlap area, center distance, and aspect ratio, ensuring the basic stability of localization. It also utilizes its non-monotonic focusing mechanism to overcome the limitations of traditional static geometric constraints. This strategy can dynamically adjust the contribution rate of samples of different qualities to the gradient, effectively mitigating training oscillations caused by outlier samples in industrial scenarios and significantly improving the sub-pixel accuracy of bounding box regression.

[0035] Specifically, at the bounding box localization level, this embodiment achieves deep coupling between geometric constraints and sample quality perception by integrating Wise-CIoU. It retains the comprehensive penalty term of CIoU for overlap area, center distance, and aspect ratio. This ensures the geometric continuity of the positioning. Its core lies in the introduction of a non-monotonic focusing coefficient. Dynamically reweight the losses: in Defined as outlier, it is used to measure the predictive quality of a sample. A sample with a low outlier (i.e., a high-quality sample) is considered a good predictor. Higher gradient weights will be assigned to promote refined convergence of the model; however, when dealing with outlier samples, i.e., when the outlier count is too large, This results in a non-monotonic descent, automatically mitigating the interference of outliers on the gradient. This dynamic allocation mechanism guides the model to converge toward a higher precision IoU threshold without increasing computational cost.

[0036] Furthermore, outlier is defined as the ratio of the IoU loss of the predicted box to the average IoU loss of all predicted boxes; When the outlier value represents a high-quality sample, the non-monotonic focusing coefficient is assigned a gradient weight that is higher than the preset threshold range. When the outlier value is represented as an outlier sample, the non-monotonic focusing coefficient produces a non-monotonic decrease to weaken gradient interference.

[0037] As a preferred implementation method, this embodiment uses a publicly available aircraft skin defect dataset. This dataset contains 3007 high-resolution images, covering five typical types of surface damage encountered during aircraft inspection: cracks, dents, missing rivets, paint-off, and scratches. To ensure the experimental results are comparable to existing public benchmarks, this experiment did not use the conventional 8:1:1 partitioning, but strictly followed the original distribution protocol of the dataset, as shown in Table 1.

[0038] Table 1 To address the challenges of uneven sample distribution and complex ambient lighting in skin defect detection tasks, this study employs a combined offline and online augmentation approach. First, basic augmentation: the original dataset already includes pre-installed offline data augmentation to ensure basic sample diversity. Second, online augmentation: during training, strategies such as random rotation, scaling, translation, and color space transformation (ColorSpaceAugmentation) are applied in real-time. This aims to simulate geometric deformation and lighting interference from the perspective of UAV inspection, significantly improving the model's ability to capture edge information of linear defects (such as scratches and cracks) and its generalization level.

[0039] Experimental Environment and Parameter Configuration: The experiments in this embodiment are implemented using Python 3.7 based on the PyTorch deep learning framework. All models were trained and evaluated on a workstation running Windows 11. Detailed hardware and software configurations are listed in Table 2.

[0040] Table 2 To ensure optimal convergence and fair comparison, all models were trained from scratch without any pre-trained weights. This strategy was adopted to rigorously validate the feature extraction capabilities of the proposed method on a specific skin defect detection task. The input image was adjusted to... The dimensions are as follows. The weights are updated using a stochastic gradient descent (SGD) optimizer. Detailed training hyperparameters are summarized in Table 3.

[0041] Table 3 To comprehensively evaluate the performance of MsDEE-Net in aircraft skin defect detection, this embodiment selects evaluation metrics from three dimensions: detection accuracy, localization accuracy, and model efficiency. Regarding detection accuracy, precision (P), recall (R), and their harmonic mean F1-Score are used, where: The positioning accuracy is centered on the mean accuracy mAP@0.5 (IoU threshold 0.5) and mAP@0.5:0.95 (average of IoU thresholds ranging from 0.5 to 0.95 with a step size of 0.05). The latter places stricter requirements on defect boundary positioning and more realistically reflects the model's spatial positioning capabilities in complex industrial environments. Model efficiency is measured by computational cost (GFLOPs), where GFLOPs represent the number of floating-point operations performed during forward inference, which is crucial for real-time detection by UAV-borne terminals.

[0042] Furthermore, this embodiment sets up an ablation experiment to systematically verify the performance contribution of each core improvement module in MsDEE-Net. This experiment uses YOLO11 as the baseline model, and adds SPD-Conv, ASE, and Wise-CIoU loss functions item by item.

[0043] Table 4 To verify the gains of each core component of MsDEE-Net in aircraft skin defect detection performance, this embodiment conducted ablation experiments item by item, and the results are shown in Table 4.

[0044] The baseline model (YOLO11) demonstrated a high recall rate (84.91%) in the skin detection task, proving its potential in candidate target discovery. However, its precision was only 71.15%. Fine-grained analysis of the prediction results revealed significant perceptual deficiencies in handling high-frequency, minute features such as cracks, and it was susceptible to artifacts caused by the highly reflective background of the skin, resulting in a high false positive rate and localization bias.

[0045] In this embodiment, by introducing the SPD-Conv structure to replace the traditional strided convolution, the model's accuracy is significantly improved from 71.15% to 78.33%, while mAP@0.5 reaches 81.37%. This mechanism, through a space-to-depth transformation strategy, reorganizes sub-pixel-level structural features to the channel dimension, effectively alleviating the information loss problem in traditional downsampling operations. Experiments demonstrate that SPD-Conv enhances the model's sensitivity to defect edge signals, thereby suppressing the generation of invalid predictions at the source of feature extraction.

[0046] In this embodiment, after integrating the ASE module, mAP@0.5 steadily increased to 83.25%. Its core gain lies in capturing the geometrically complex Dent category and edge-blurred defects. This improvement is mainly attributed to the high-frequency signal compensation provided by the EdgeBoost internal ASE module, and the filtering effect of the Dual Domain Selection (DSM) mechanism. As a selective information bottleneck, DSM adaptively decouples target features from background noise, accurately locking onto discriminative semantics highly relevant to the detection task even in environments with strong interference.

[0047] Finally, after introducing the Wise-CIoU loss function, MsDEE-Net achieved optimal overall performance. The mAP@0.5:0.95 peaked at 51.80%, an improvement of 4.31 percentage points compared to the ASE stage. Particularly noteworthy is that for the most challenging Crack category, its mAP@0.5 ultimately reached 69.78%, far exceeding the baseline model's 60.87% (see the category comparison analysis below for details). This verifies the superiority of the non-monotonic focusing mechanism in handling uneven sample quality distribution, successfully guiding the model towards sub-pixel-level optimal bounding boxes.

[0048] Furthermore, to comprehensively evaluate the performance of the proposed MsDEE-Net algorithm in complex defect detection tasks, this embodiment also quantitatively compares it with current mainstream object detection algorithms under the same experimental settings. The comparison models include the classic two-stage detector Faster R-CNN (ResNet50) and the advanced one-stage detector series: YOLOv8s, YOLOv10s, YOLO11s, and YOLO12s. Table 5 shows the comprehensive performance of each model on key metrics such as Precision, Recall, mAP@0.5, mAP@0.5:0.95, GFLOPs, and FPS. Tables 6 to 11 further detail the specific detection results for each category. The specific detection results of Faster R-CNN in each category are shown in Table 6, the specific detection results of YOLOv8s in each category are shown in Table 7, the specific detection results of YOLOv10s in each category are shown in Table 8, the specific detection results of YOLO11s in each category are shown in Table 9, the specific detection results of YOLO12s in each category are shown in Table 10, and the specific detection results of MsDEE-Net in each category are shown in Table 11.

[0049] Table 5 In terms of overall detection performance, MsDEE-Net achieved top-tier performance across most key metrics. Specifically, MsDEE-Net achieved mAP@0.5 of 87.47% and mAP@0.5:0.95 of 51.80%, representing improvements of 7.15% and 2.43% respectively compared to the strongest baseline model, YOLO11s. This indicates that the improved network structure significantly enhances feature extraction and multi-scale fusion capabilities, enabling more accurate coverage of defect targets. Notably, while maintaining a high recall rate (Recall=83.82%), MsDEE-Net significantly improved detection precision (Precision=82.54%), an improvement of 11.39% compared to YOLO11s' 71.15%. This result strongly demonstrates that MsDEE-Net can effectively suppress background interference and significantly reduce false positives, which is of great significance for industrial quality inspection scenarios aiming for zero defect outflow.

[0050] In comparison with the classic two-stage algorithm Faster R-CNN, MsDEE-Net demonstrated an overwhelming advantage. Although Faster R-CNN uses a deep ResNet50 backbone network, its mAP@0.5 is only 65.47%, and its computational cost is as high as 948.18 GFLOPs. This significant difference in computational cost stems from the fundamental difference in their architectures: Faster R-CNN's GFLOPs are significantly higher than those of the YOLO series, mainly because in its two-stage architecture, the large number of proposals (candidate regions) generated by the RPN need to be calculated one by one through the ROIHead, and this embodiment reports the total inference computational cost including all proposal processing. In contrast, the YOLO series, as a one-stage detector, mainly concentrates its computational cost on the convolutional operations of the backbone and neck, and typically reports MACs. In comparison, MsDEE-Net reduces computation by approximately 98.8% (only 11.5 GFLOPs), while increasing inference speed by 68% (50.7 FPS vs 85.23 FPS) and accuracy by 22%, fully demonstrating the dual superiority of the method in this embodiment in terms of efficiency and accuracy.

[0051] A thorough analysis of the detection results for each defect category reveals that MsDEE-Net has achieved a breakthrough in detecting extremely difficult-to-detect minute defects. Taking the "Scratch" category as an example, due to its weak features and varied shapes, YOLOv8s and YOLO12s achieve mAP@0.5 of only 60.61% and 55.75% respectively, with extremely low precision (e.g., YOLO10s only 25.12%), indicating that existing models are prone to misclassifying background textures as scratches. In contrast, MsDEE-Net achieves 99.50% mAP@0.5 and 100% precision in this category, with an F1-Score as high as 0.9389. This improvement is mainly attributed to the effective preservation of low-resolution features by the SPD-Conv module introduced in this embodiment, and the keen capture of subtle features by the multi-head attention mechanism. In addition, MsDEE-Net also maintained its leading position in categories such as "Dent" and "Paint-off", demonstrating the model's strong robustness in dealing with multi-scale and multi-morphological defects.

[0052] Finally, regarding inference efficiency, although MsDEE-Net's GFLOPs are slightly higher than the YOLO series baseline due to the introduction of additional feature enhancement modules, resulting in a slight decrease in FPS from 98.18 (YOLO11s) to 85.23, this speed still far exceeds the 30 FPS standard typically required for industrial real-time inspection. Moreover, this increased computational cost is traded for a significant 7.15% improvement in accuracy and an exceptional ability to capture minute defects, making it highly valuable for practical applications. In conclusion, MsDEE-Net successfully breaks through the performance bottleneck of existing lightweight models in complex defect detection, providing an excellent solution that combines high accuracy and real-time performance.

[0053] Table 6 Table 7 Table 8 Table 9 Table 10 Table 11 To intuitively verify the effectiveness of MsDEE-Net, this embodiment selects aircraft skin surface images with complex backgrounds for inference testing. In particular, for shallow scratches on the skin surface, MsDEE-Net demonstrates stronger continuity detection capability than the baseline, avoiding the phenomenon of identifying a long scratch break as multiple fragmented targets.

[0054] like Figure 4 and Figure 5 As shown in the visualization comparison, MsDEE-Net demonstrates a significant technological leap in robustness against extremely small targets. In the detection results of the baseline model, limited by the pixel-level detail loss caused by the traditional convolutional downsampling process, the model exhibits a clear risk of missing detections of minute scratches and early micro-cracks, resulting in the annihilation of many key defect features in the deep network. In contrast, MsDEE-Net, with its introduced SPD-Conv lossless downsampling structure, effectively preserves the spatial topological information of the original image, achieving accurate feature triggering and capture even when faced with damage of extremely small scale and indistinguishable to the naked eye.

[0055] Further observation of the prediction results reveals that MsDEE-Net achieves a dual leap in both confidence score and localization accuracy. The baseline model's predictions are relatively scattered in the medium-to-high confidence region and are prone to localization shifts and false detections in complex defect morphologies. The Precision-Recall curves show a rapid decrease in Precision in the medium-to-high Recall range, reflecting the model's difficulty in maintaining accuracy while preserving detection completeness. Thanks to the selective enhancement of edge features by the ASE module and the optimization of bounding box regression using the Wise-CIoU loss function, MsDEE-Net's PR curves across all categories shift generally to the upper right, maintaining high Precision, especially in the high Recall region. This improvement is not only reflected in numerical gains but also in the closer fit of the predicted bounding box to the defect's geometric contour, significantly reducing localization errors caused by background texture interference.

[0056] Furthermore, MsDEE-Net demonstrates superior engineering anti-interference capabilities against common issues on aircraft skin surfaces, such as rivet interference, spectral reflection, and non-uniform coating textures. Thanks to the integrated energy filtering function of the DSM, the system can filter out redundant background noise in both the spatial and frequency domains. Unlike baseline models, which are prone to misjudgments or severe fluctuations in predicted bounding boxes when faced with texture noise, MsDEE-Net's visualization detection results exhibit extremely high cleanliness and output stability. This robust detection performance under complex conditions demonstrates the algorithm's strong practicality and widespread application value in real-world aviation maintenance scenarios.

[0057] In summary, the experimental results of this embodiment fully verify the advancement and effectiveness of the MsDEE-Net architecture in aircraft skin defect detection. Compared with current mainstream YOLO series algorithms, MsDEE-Net not only achieves comprehensive superiority in core accuracy metrics such as mAP@0.5 and mAP@0.5:0.95, but also demonstrates outstanding characterization capabilities for extremely subtle defects such as cracks and scratches.

[0058] This breakthrough performance is primarily attributed to the ASE module's precise compensation for high-frequency edge signals and the SPD-Conv's lossless preservation of sub-pixel features. Together, these two technologies address the industry pain point of feature loss in complex backgrounds. Furthermore, qualitative analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps further confirms that the model, driven by DSM, can significantly suppress noise and achieve high focus on damaged areas. Ultimately, MsDEE-Net maintains an industrial-grade real-time inference speed of 85.23 while preserving extremely high detection accuracy, perfectly meeting the real-time and accuracy requirements of UAV-borne terminals in complex inspection environments. This provides an efficient and reliable deep learning solution for the automated detection of aircraft surface damage.

[0059] This embodiment addresses the urgent need for structural integrity monitoring and flight safety assurance of aircraft in complex service environments. First, it introduces a lossless feature reconstruction mechanism (SPD-Conv), which retains key sub-pixel information while reducing feature scale through spatial-depth transformation, fundamentally solving the problem of information annihilation during downsampling of minute defects. Second, it designs an adaptive selective edge enhancement (ASE) module, which effectively compensates for the weakening of subtle signals such as cracks and scratches by traditional convolution using high-frequency differential logic. Then, it integrates a dual-domain selection mechanism (DSM), which simultaneously focuses discriminative features in spatial and frequency dimensions, significantly suppressing interference from metal skin reflections and complex background noise. Finally, it introduces a Wise-CIoU loss function with an integrated non-monotonic focusing mechanism, which achieves deep coupling between geometric constraints and sample quality perception by dynamically adjusting the gradient weights of samples of different quality, significantly improving the localization accuracy and convergence stability of bounding box regression.

[0060] Extensive comparative experiments demonstrate that MsDEE-Net achieves an mAP@0.5 of 87.47% on the ASDD dataset, significantly outperforming mainstream benchmark algorithms such as YOLO11s, especially in the representation of extremely subtle damages like Cracks and Scratch. Heatmap analysis further confirms the model's excellent interpretability and target focusing capabilities. While maintaining extremely high detection accuracy, its inference speed of 85.23 FPS fully meets the real-time requirements of UAV-based terminals in practical inspection scenarios, achieving an ideal industrial-grade balance between performance and efficiency.

[0061] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting surface defects on aircraft, characterized in that, include: Based on the acquired aircraft skin surface image, the first feature map is obtained by lossless feature reconstruction through a spatial-depth convolution module. Based on the first feature map, a second feature map is obtained by performing multi-scale feature extraction and dual-domain selection enhancement through an adaptive selective edge enhancement module. Based on the second feature map, the bounding box regression is dynamically optimized by the Wise-CIoU loss function with integrated non-monotonic focusing mechanism to obtain the aircraft surface defect detection results.

2. The method according to claim 1, characterized in that, The process of obtaining the first feature map includes: The aircraft skin surface image is subjected to spatial-depth transformation, and multiple sub-feature maps are generated through periodic sampling; Multiple sub-feature maps are concatenated along the channel dimension to obtain an intermediate feature map with reduced spatial resolution and expanded channel number. The first feature map is obtained by performing cross-channel feature interaction and channel number mapping on the intermediate feature map through a convolutional layer with a stride of 1.

3. The method according to claim 1, characterized in that, The process of obtaining the second feature map includes: The first feature map is input into the adaptive selective edge enhancement module, and multi-scale features and local detail features are extracted through parallel processing of multi-scale perception path and local perception path, respectively. The multi-scale features are subjected to edge enhancement processing to obtain the edge-enhanced multi-scale features; The local detail features and the edge-enhanced multi-scale features are concatenated along the channel dimension to obtain the fused features; The fused features are globally evaluated and filtered in the spatial and frequency domains using a dual-domain selection mechanism to obtain purified features. The purified features are subjected to channel dimensionality reduction and recombination to obtain the second feature map.

4. The method according to claim 3, characterized in that, The process of edge enhancement processing of the multi-scale features includes: Average pooling is performed on the input feature map to extract low-frequency background information; Subtract the low-frequency background information from the input feature map to extract high-frequency information containing edge details; The high-frequency information is passed through a convolutional layer with a Sigmoid activation function to generate an edge attention weight map; The edge attention weight map is superimposed back onto the input feature map to complete the adaptive activation and enhancement of edge features.

5. The method according to claim 3, characterized in that, The construction process of the multi-scale sensing path includes: By using multiple parallel adaptive average pooling layers, local semantic signals at different spatial scales are captured to obtain multi-scale pooling features. The multi-scale pooling features are convolved and then upsampled to restore the original resolution to obtain the multi-scale features.

6. The method according to claim 1, characterized in that, The Wise-CIoU loss function is obtained by multiplying the CIoU loss term by the non-monotonic focusing coefficient; The non-monotonic focusing coefficient is dynamically determined based on the outlier degree between the predicted box and the true box.

7. The method according to claim 6, characterized in that, The outlier is defined as the ratio of the IoU loss of the predicted box to the average IoU loss of all predicted boxes. When the outlier value represents a high-quality sample, the non-monotonic focusing coefficient is assigned a gradient weight that is higher than a preset threshold range. When the outlier value represents an outlier sample, the non-monotonic focusing coefficient decreases non-monotonically to weaken gradient interference.

Images