Aircraft skin defect detection method and system based on double statistical sparse experts and bar depth convolution

Abstract

The application discloses an aircraft skin defect detection method and system based on double statistical sparse experts and bar depth convolution, and relates to the technical field of aircraft skin defect detection. The method sets a double statistical sparse mixed expert module in the backbone network and / or neck network of a detection network, and sets a bar depth convolution module in the high-resolution feature path of the neck network, realizes adaptive feature extraction of multi-form defects through a double statistical descriptor and a sparse expert selection mechanism, and enhances the directional context expression of slender defects through bar depth convolution. The application solves the problems of high micro low-contrast defect omission rate, incomplete slender defect detection, lack of unified adaptive framework for multiple types of defects, and difficulty in meeting the edge end lightweight real-time deployment, can improve the detection capability of skin defects such as cracks, pits, rivet head loss, paint loss and scratches, and is suitable for unmanned aerial vehicle inspection, handheld detection equipment and hangar fixed visual detection systems.

Description

Technical Field

[0001] This invention relates to the field of aircraft skin defect detection technology, and in particular to a lightweight aircraft skin defect detection method and system based on dual statistical sparse experts and strip depth convolution. Background Technology

[0002] Aircraft skin is a critical component ensuring the structural safety and flight performance of an aircraft, and its surface quality directly affects flight safety. During manufacturing, maintenance, and long-term service, aircraft skin is prone to various defects such as cracks, dents, missing rivet heads, paint peeling, and scratches. If these defects are not detected and repaired in a timely manner, they may further expand and weaken the structural strength of the skin, and in severe cases, endanger flight safety.

[0003] Currently, aviation companies generally use manual visual inspection to detect defects in aircraft skin. This method relies on the experience and subjective judgment of the inspectors, resulting in low inspection efficiency, high missed detection rate, and poor consistency in judgment, making it difficult to meet the accuracy requirements for aircraft skin defect detection.

[0004] In recent years, with the development of deep learning technology, deep learning-based object detection algorithms, especially the YOLO series, have been attempted for skin defect detection. For example, YOLO26, as a relatively new end-to-end object detection model, has improved detection efficiency. However, when directly applying general object detection networks to aircraft skin defect detection, the following technical problems still exist:

[0005] First, the aircraft skin surface has complex backgrounds such as metallic textures and uneven lighting. Small defects such as pits and paint peeling occupy very few pixels in the image and have low contrast. Their weak features are easily submerged in the background texture, and existing detection networks have difficulty effectively extracting them, resulting in a high rate of missed detection of small defects.

[0006] Secondly, defects such as cracks and scratches appear as long and thin lines with a very high aspect ratio in images. Conventional anchor point allocation mechanisms and feature extraction operators are difficult to effectively cover their global structure, resulting in incomplete detection results and a high rate of missed detections.

[0007] Although some improvement schemes have been developed, such as using dynamic snake convolution or introducing attention modules, most of them are designed for single defect morphology and lack a unified adaptive feature extraction framework for multiple types of defects. At the same time, such improvement schemes are usually accompanied by large computational overhead and number of parameters, which makes it difficult to meet the needs of real-time detection in edge computing scenarios with limited computing power.

[0008] Therefore, there is an urgent need for an aircraft skin defect detection solution that can not only enhance the feature representation capabilities of small and slender defects, but also perform unified adaptive processing on multiple types of defects, while maintaining a lightweight design and being suitable for real-time deployment on edge devices. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a lightweight aircraft skin defect detection method and system based on dual statistical sparse experts and strip deep convolution. By setting dual statistical sparse hybrid expert modules in the backbone network and / or neck network, and setting strip deep convolution modules in the high-resolution feature paths of the neck network, this invention solves the technical problems of high defect false negative rates, lack of a unified adaptive framework for multiple defect types, and difficulty in meeting the requirements of lightweight real-time deployment of edge devices in existing technologies.

[0010] According to an embodiment of the present invention, a method for detecting aircraft skin defects based on dual statistical sparse experts and stripe depth convolution includes the following steps:

[0011] Obtain aircraft skin images;

[0012] The aircraft skin image is input into a YOLO26-based detection network, which includes a backbone network, a neck network, and a detection head.

[0013] A dual statistical sparse hybrid expert module is set in the backbone network and / or the neck network. The expert pool of the dual statistical sparse hybrid expert module includes at least two of the following: point defect experts, slender defect experts, fuzzy boundary experts, and identity mapping experts. The dual statistical sparse hybrid expert module calculates the spatial mean and spatial peak value of the input feature map, constructs a dual statistical descriptor based on the spatial mean and spatial peak value, inputs the dual statistical descriptor into a lightweight router to obtain the routing score of each expert, performs normalization processing and Top-k sparse selection on the routing score to obtain sparse expert weights, and sums the outputs of at least one selected expert branch in the expert pool according to the sparse expert weights to obtain the output feature map of the dual statistical sparse hybrid expert module.

[0014] A strip deep convolution module is set in the high-resolution feature path of the neck network. The strip deep convolution module performs horizontal strip deep convolution and vertical strip deep convolution in parallel, and the output of the horizontal strip deep convolution is fused with the output of the vertical strip deep convolution to obtain an output feature map that enhances the directional contextual expression of slender defects.

[0015] The detection network outputs the detection results of aircraft skin defects through the detection head, and the detection results include the defect type and defect location.

[0016] Preferably, constructing the dual statistical descriptor based on the spatial mean and the spatial peak value includes: concatenating the spatial mean and the peak value deviation to construct the dual statistical descriptor, wherein the spatial mean... and spatial peak Calculate using the following formula:

[0017]

[0018]

[0019] in, This represents the input feature map, where C, H, and W represent the number of channels, height, and width of the input feature map, respectively. These are the channel index, height index, and width index, respectively.

[0020] The spatial mean and peak deviation are concatenated to construct the dual statistical descriptor. :

[0021]

[0022] The lightweight router is a two-layer multilayer perceptron (MLP), comprising a linear transformation, an activation function, and a linear output connected in sequence, used to nonlinearly map the dual statistical descriptor S to routing scores from each expert. , ,in, This represents the total number of experts in the expert pool.

[0023] Preferably, the step of normalizing the routing score and obtaining sparse expert weights through Top-k sparse selection specifically includes: performing Softmax normalization on the routing score to obtain normalized weights, selecting the top two experts with the largest normalized weights, resetting the weights of the unselected experts to zero, and renormalizing the non-zero weights to obtain sparse expert weights.

[0024] Preferably, the expert pool includes:

[0025] Point defect expert, for 3×3 depthwise separable convolution;

[0026] Slender Defect Expert, for strip-shaped depthwise convolutional modules;

[0027] Fuzzy boundary expert, which is a dilated depthwise convolution with a dilation rate of 2, and connected to a 1×1 convolution;

[0028] An expert in identity mapping, specializing in identity mapping.

[0029] Preferably, in the dual statistical sparse hybrid expert module, the input feature map X is reduced in dimensionality by a shared 1×1 preprocessing convolution to obtain the intermediate feature Z. Each expert in the expert pool processes the intermediate feature Z. The outputs of the selected expert branches are weighted and summed according to the sparse expert weights, then increased in dimensionality by a 1×1 convolution, and finally added to the result obtained by quick connection of the input feature map X to serve as the output feature map of the dual statistical sparse hybrid expert module.

[0030] Preferably, the horizontal bar depth convolution is a depth convolution with a kernel size of 1×k, and the vertical bar depth convolution is a depth convolution with a kernel size of k×1. The number of groups in both the horizontal and vertical bar depth convolutions is equal to the number of channels C of the input feature map. The first processed feature obtained by batch normalization and SiLU activation of the horizontal bar depth convolution is added to the second processed feature obtained by batch normalization and SiLU activation of the vertical bar depth convolution, and then output after depthwise separable convolution.

[0031] Preferably, the backbone network sequentially includes a first convolution, a second convolution, a first feature fusion module, a third convolution, a second feature fusion module, a fourth convolution, a third feature fusion module, a fifth convolution, an SPPF module, and a fourth feature fusion module; wherein, the first feature fusion module, the second feature fusion module, and the fourth feature fusion module are C2f modules, the third feature fusion module is a feature fusion module with 512 output channels, and the original C3k2 module or C2PSA module at the position of the third feature fusion module is replaced by the dual statistical sparse hybrid expert module; the SPPF module is an SPPF module with shortcut links.

[0032] Preferably, the neck network adopts an FPN-PAN structure, including an FPN upsampling path and a PAN downsampling path. The neck network outputs P3 features, P4 features and P5 features, wherein P3 features, P4 features and P5 features are the highest resolution features, medium resolution features and lowest resolution features, respectively, and correspond to the detection of small targets, medium targets and large targets, respectively.

[0033] The high-resolution feature path is the highest resolution P3 generation path of the neck network, and the strip-shaped depth convolution module is set in the P3 generation path.

[0034] At least one C2f module in the neck network is replaced with the dual statistical sparse hybrid expert module; wherein, the output feature map of the dual statistical sparse hybrid expert module is used as a feature map in the backbone network and / or the neck network to participate in feature extraction and feature fusion, and the output feature map of the strip deep convolution module is used as a feature map in the high-resolution feature path of the neck network to participate in feature fusion;

[0035] The neck network outputs the processed feature map to the detection head;

[0036] The detection head includes three detection branches, which perform defect detection on features P3, P4 and P5 respectively, and output the detection results of aircraft skin defects.

[0037] Preferably, the detection network is a YOLO26-nano model, and the detection network adopts an end-to-end NMS-Free dual-head architecture; the defect types include cracks, dents, missing rivet heads, paint peeling, and scratches.

[0038] On the other hand, according to embodiments of the present invention, an aircraft skin defect detection system based on dual statistical sparse experts and stripe depth convolution is also provided, comprising:

[0039] The image acquisition module is used to acquire images of the aircraft skin.

[0040] The defect detection module is used to perform defect detection on the aircraft skin image using the aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in any of the above technical solutions, and output the defect type and defect location.

[0041] The inspection report generation module is used to generate an inspection report based on the defect type and defect location.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention uses a dual statistical sparse hybrid expert module to perform statistical analysis on input features, constructs a routing descriptor that takes into account both the overall distribution of features and the local significant response, and dynamically selects the most suitable processing operator for different features. This overcomes the shortcomings of existing methods that are not sensitive to weak defect feature responses and are prone to missing small defects, and effectively improves the detection accuracy of small low-contrast defects such as small pits and small areas of paint peeling.

[0044] This invention captures long-distance strip-shaped contextual information with extremely low parameter cost through a strip-shaped deep convolution module, obtaining a receptive field that matches the morphology of defects with large aspect ratios such as cracks and scratches. This overcomes the problem that conventional feature extraction methods are unable to cover the global structure of slender defects and result in incomplete detection, effectively improving the detection capability of slender defects.

[0045] This invention utilizes a sparse expert selection mechanism composed of complementary experts such as point-like, elongated, and fuzzy boundary experts to dynamically match the most suitable feature extraction operator for defects of different shapes. This overcomes the shortcomings of existing improved schemes that are only designed for a single defect shape and lack a unified adaptive feature extraction framework for multiple types of defects. At the same time, the sparse selection mechanism only activates a small number of expert branches, and the overall parameter count can be controlled below 3M, which can meet the deployment requirements of real-time detection in edge devices and scenarios with limited computing power, such as UAV inspection, handheld inspection equipment, and fixed vision inspection systems in hangars.

[0046] Therefore, this invention improves the detection capability of various types of aircraft skin defects, such as cracks, dents, missing rivet heads, paint peeling, and scratches, while maintaining the lightweight nature of the model. Attached Figure Description

[0047] Figure 1 The flowchart of the aircraft skin defect detection method based on dual statistical sparse experts and strip depth convolution provided by the present invention is shown below.

[0048] Figure 2 This is a diagram illustrating the overall architecture of the detection network provided by the present invention.

[0049] Figure 3 A schematic diagram of the structure of the DSMoELite module, a dual statistical sparse hybrid expert provided by the present invention;

[0050] Figure 4 A schematic diagram of the StripDWConv module provided by the present invention;

[0051] Figure 5 This is a schematic diagram of the structure of the aircraft skin defect detection system based on dual statistical sparse experts and strip depth convolution provided by the present invention. Detailed Implementation

[0052] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. The specific values ​​and strategies described in the embodiments are exemplary, and those skilled in the art can make adaptive adjustments according to actual needs.

[0053] like Figure 1 As shown, this embodiment of the invention proposes a method for detecting defects in lightweight aircraft skin based on dual statistical sparse experts and stripe depth convolution, including the following steps:

[0054] Obtain aircraft skin images;

[0055] The aircraft skin image is input into a YOLO26-based detection network, which includes a backbone network, a neck network, and a detection head.

[0056] A dual statistical sparse hybrid expert module is set in the backbone network and / or the neck network. The expert pool of the dual statistical sparse hybrid expert module includes at least two of the following: point defect experts, slender defect experts, fuzzy boundary experts, and identity mapping experts. The dual statistical sparse hybrid expert module calculates the spatial mean and spatial peak value of the input feature map, constructs a dual statistical descriptor based on the spatial mean and spatial peak value, inputs the dual statistical descriptor into a lightweight router to obtain the routing score of each expert, performs normalization processing and Top-k sparse selection on the routing score to obtain sparse expert weights, and sums the outputs of at least one selected expert branch in the expert pool according to the sparse expert weights to obtain the output feature map of the dual statistical sparse hybrid expert module.

[0057] A strip deep convolution module is set in the high-resolution feature path of the neck network. The strip deep convolution module performs horizontal strip deep convolution and vertical strip deep convolution in parallel, and the output of the horizontal strip deep convolution is fused with the output of the vertical strip deep convolution to obtain an output feature map that enhances the directional contextual expression of slender defects.

[0058] The detection network outputs the detection results of aircraft skin defects through the detection head, and the detection results include the defect type and defect location.

[0059] It should be noted that the dual statistical sparse hybrid expert module is the DSMoELite module, and the strip deep convolution module is the StripDWConv module.

[0060] In one specific embodiment, YOLO26-nano is used as the baseline model. The model is trained and tested on a dataset of aircraft skin defects with labeled images. Five types of defects are detected: crack, dent, missing rivet head, paint-off, and scratch. Before training, images of the aircraft skin surface are collected and uniformly scaled to 640×640 pixels and normalized.

[0061] In this embodiment, the overall architecture of the detection network is as follows: Figure 2As shown, the network follows the YOLO26 standard architecture, consisting of a backbone network, a neck network, and a detection head. The backbone network sequentially includes the following modules: a first convolutional layer (Conv(3→64, s=2)), a second convolutional layer (Conv(64→128, s=2)), a first feature fusion module (C2f module, 128→256), a third convolutional layer (Conv(256→256, s=2)), a second feature fusion module (C2f module, 256→512), a fourth convolutional layer (Conv(512→512, s=2)), a third feature fusion module, a fifth convolutional layer (Conv(512→1024, s=2)), an SPPF module (with shortcut connections), and a fourth feature fusion module (C2f module, 1024). The first, second, and fourth feature fusion modules are all C2f modules, while the third feature fusion module has 512 output channels.

[0062] This embodiment replaces the original C3k2 module (YOLO series feature fusion module) or C2PSA module (feature fusion module integrating pyramid splitting attention) at this position with a dual statistical sparse hybrid expert module (i.e., the replacement module can be either C3k2 or C2PSA). This position is located at the convergence node of deep features in the backbone network, simultaneously receiving shallow detail information and deep semantic information. Introducing the dual statistical sparse hybrid expert module here allows the network to dynamically select the most suitable convolution operator based on the statistical characteristics of the input features themselves with minimal replacement cost, achieving adaptive feature extraction for multi-morphological defects. The remaining feature fusion modules continue to use the C2f module, which maximizes the inheritance of the YOLO26 pre-trained structure, which is beneficial for training convergence and parameter lightweighting. It should be noted that setting the dual statistical sparse hybrid expert module in the backbone network or neck network means replacing at least one feature fusion module in the backbone network or neck network with this module. The feature fusion module can be the C3k2 module, the C2PSA module, or other convolutional feature fusion modules.

[0063] Following the backbone network, the neck network undertakes the task of multi-scale feature fusion. The scale of defects in aircraft skin is extremely wide, ranging from pixel-level small pits and small areas of paint peeling to cracks and scratches covering large areas. Features of a single scale are difficult to meet the detection requirements of defects of different scales at the same time. Therefore, the neck network in this embodiment adopts an FPN-PAN structure. The FPN upsampling path upsamples and splices shallow features step by step to retain the detailed information of small targets, and the PAN downsampling path downsamples and splices deep features step by step to enhance the semantic information of large targets, thereby achieving full fusion of multi-scale features.

[0064] Specifically, the FPN upsampling path is as follows: The first upsampling receives the output features of the fourth feature fusion module C2f in the backbone network, concatenates them with the output features of the DSMoELite module in the backbone network at the first splicing node, and then sends them to the first DSMoELite module; The output features of the first DSMoELite module are upsampled a second time and then concatenated with the output features of the second feature fusion module C2f in the backbone network at the second splicing node. The concatenated features first pass through the MicroDetailEnhancer module and then are sent to the second DSMoELite module; The output features of the second DSMoELite module are upsampled a third time and then concatenated with the output features of the first feature fusion module in the backbone network at the third splicing node, and then sent to the StripDWConv module. After processing by the top C2f module, the P3 features (128 channels) are obtained. After the second stitching node, a MicroDetaiEnhancer module is set up. The stitching features contain multi-scale semantic information and spatial detail information. The response of small defects may still be masked by the strong background texture response after stitching. By first performing microdetail enhancement processing and then sending it to the subsequent module, the detailed features of small defects can be avoided from being diluted in the further feature processing, thereby improving the expression of detailed features of small low-contrast defects such as small pits and small areas of paint peeling.

[0065] It should be noted that the MicroDetailEnhancer module is used to enhance the feature response of minute defects in the feature map. Specifically, the MicroDetailEnhancer module includes a multi-branch parallel feature extraction unit, a feature recalibration unit, and a feature fusion output unit. The multi-branch parallel feature extraction unit is used to extract multi-dimensional high-frequency features from the input feature map. It includes a regular convolution branch and at least one differential convolution branch; the differential convolution branch is used to calculate the difference between each pixel in the input feature map and its neighboring pixels to extract high-frequency detail features; optionally, the differential convolution branch includes one or more of central differential convolution, angular differential convolution, and horizontal differential convolution; the feature recalibration unit is connected to the multi-branch parallel feature extraction unit and is used to filter and enhance the extracted multi-dimensional high-frequency features through an attention mechanism to suppress background noise and enhance the response of minute defects.

[0066] Tiny defects correspond to high-frequency components in the image frequency domain. The low-pass filtering characteristic of conventional convolution operations causes these components to attenuate and be lost during layer-by-layer propagation. Therefore, the MicroDetaiEnhancer module actively extracts high-frequency details through multi-directional differential convolution, intelligently filters effective details and noise through a channel attention mechanism, and then injects the enhanced details into the feature map without loss through residual connections, thereby effectively improving the network's ability to detect tiny low-contrast defects.

[0067] In the PAN downsampling path, the P3 feature, after convolution, is concatenated with the output of the first DSMoELite module at the fourth concatenation node, and then processed by the C2f module to generate the P4 feature. The P4 feature, after convolution, is concatenated with the output of the second DSMoELite module at the fifth concatenation node, and then processed by the C2f module to generate the P5 feature. Simultaneously, some standard C2f modules in the neck network are replaced with dual statistical sparse hybrid expert modules. The neck network undertakes the multi-scale feature fusion task. Since the morphological manifestations of defects differ across different scales, the dual statistical sparse hybrid expert modules can adaptively select fusion operators based on the statistical characteristics of the features themselves, improving the representational ability of multi-morphological defects in the feature fusion stage.

[0068] After the aforementioned feature extraction and fusion, the detection head comprises three detection branches, which respectively perform defect detection on features P3, P4, and P5, outputting the defect type and location. The reason for employing a three-branch hierarchical detection approach is that features P3 have the highest resolution, corresponding to small targets (such as tiny pits or small areas of paint peeling), primarily carrying detailed information about small targets; features P4 correspond to medium-sized targets (medium-scale defects); and features P5 have the lowest resolution, corresponding to large targets (such as large areas of paint peeling or through-cracks), primarily carrying semantic information about large targets. The three branches match the scale distribution of the defects, avoiding missed detections at extreme scales caused by single-branch detection. The detection head adopts the YOLO26 end-to-end NMS-Free dual-head architecture, thereby reducing inference latency and meeting the requirements of real-time detection.

[0069] The following describes the specific implementation of the dual-statistical sparse hybrid expert module, whose structure is as follows: Figure 3 As shown, this module dynamically selects the most suitable convolution operator for the current input features to adapt to defects of different shapes and scales. Addressing the problem of ambiguous routing decisions caused by traditional Hybrid Expert (MoE) routing relying solely on a single statistic (such as global average), this module employs a dual-statistic routing mechanism. Specifically, the input feature map... First, the intermediate feature Z is obtained through 1×1 preprocessing convolutional dimensionality reduction, where C, H, and W are the number of channels, height, and width of the input feature map, respectively. All expert branches in the expert pool use the same 1×1 convolutional layer to perform channel dimensionality reduction on the input feature map X, generating an intermediate feature Z. The intermediate feature Z is then input to each selected expert branch. Through this shared design, it is ensured that each expert branch processes the same intermediate feature after uniform transformation, maintaining consistency between routing decisions and expert processing. Furthermore, it significantly reduces the number of parameters in the module, which is beneficial for lightweight deployment on the edge.

[0070] Meanwhile, the spatial mean of the input feature map X is calculated using global average pooling and global max pooling, respectively. and spatial peak The calculation formula is:

[0071]

[0072]

[0073] in, This represents the input feature map, where C, H, and W represent the number of channels, height, and width of the input feature map, respectively. These are the channel index, height index, and width index, respectively.

[0074] The spatial mean and peak deviation are concatenated to construct the dual statistical descriptor. :

[0075]

[0076] The peak deviation is the difference between the spatial peak value and the spatial mean value.

[0077] The spatial mean reflects the global distribution of the overall energy of the channel, while the peak bias amplifies the prominence of local salience responses within the channel relative to the overall level. The two complement each other perfectly: when relying solely on the global mean, the weak local responses of small defects in the metallic texture background will be diluted by the mean, and the routing decision will be insensitive to small defects; with the addition of the peak bias, the routing decision can simultaneously perceive the global distribution and local salience, thereby enabling more accurate operator selection for small low-contrast defects.

[0078] Subsequently, the dual statistical descriptors S are input into the lightweight router to obtain the routing scores of each expert. The lightweight router is a two-layer multilayer perceptron (MLP), exemplarily employing a Linear+SiLU+Linear structure, containing sequentially connected linear transformations, activation functions, and linear outputs, used to nonlinearly map the dual statistical descriptors S to the routing scores of each expert. , The total number of experts in the expert pool. After obtaining the routing score, it is first Softmax normalized to obtain normalized weights, and then Top-k (preferably k=2 in this invention) sparse selection is performed: the top two experts with the largest normalized weights are selected, the weights of the unselected experts are reset to zero, and the non-zero weights are renormalized to obtain sparse expert weights; the renormalization divides the weights of the experts retained after Top-selection by their sum, so that the sum of the weights of the selected experts is equal to 1 again, thereby ensuring the numerical stability of the subsequent weighted fusion of expert outputs. The reason for using Top-2 sparse selection and renormalization is that if all experts participate in the calculation, the computational cost will increase linearly with the number of experts, which violates the lightweight goal; only activating the two expert branches with the largest weights can significantly reduce the computational cost while maintaining the feature extraction capability, and renormalizing the non-zero weights can ensure the stability of the energy scale of the output features and avoid numerical fluctuations caused by sparse activation.

[0079] Based on the aforementioned sparse expert weights, the four complementary experts in the expert pool process the intermediate feature Z respectively:

[0080] Expert 0 is a point defect expert, using a 3×3 depthwise separable convolution (DWConvBNAct), which is good at extracting point defect features and is suitable for small, low-contrast defects such as small pits and small areas of paint peeling.

[0081] Expert 1 is an expert in slender defects, using the StripDWConv module, which is specifically designed to extract slender defect features such as cracks and scratches;

[0082] Expert 2 is a fuzzy boundary expert, which is a dilated depthwise convolution with a dilation rate of 2 followed by 1×1 convolutions, used to expand the receptive field and adapt to fuzzy boundaries or large defects.

[0083] Expert 3 is an identity mapping expert, which is an identity mapping (nn.Identity) used to handle simple background features and reduce computation.

[0084] Aircraft skin defects exhibit diverse morphologies, and a single convolutional operator cannot simultaneously adapt to defects of different shapes, such as point-like, elongated, and blurred boundaries. Four experts are assigned to different defect morphologies, and an identity mapping expert covers a simple background region. This allows the routing mechanism to match the most suitable operator for different defect morphologies based on feature statistical characteristics, achieving a unified adaptive feature extraction framework. After each expert's processing, the outputs of the selected expert branches are weighted and summed according to sparse expert weights, then subjected to 1×1 convolution for dimensionality increase. This summation is then added to the result obtained by shortcut connection of the input feature map X, serving as the output feature map of the dual statistical sparse hybrid expert module. The use of 1×1 convolution for dimensionality increase and adding the result to the shortcut connection serves two purposes: firstly, it restores the weighted summation result to a value consistent with the input feature map. Figure 1The number of channels is consistent, making it easy to add them; whether the target number of channels for 1×1 convolution dimensionality enhancement is restored to the number of channels of the input feature map X; on the other hand, the shortcut connection retains the original information of the input features, forming a residual learning structure, which is beneficial to gradient backpropagation and training stability, while avoiding damage to the trained features after the module is introduced.

[0085] In summary, in the dual-statistical sparse hybrid expert module, the spatial mean reflects the overall energy distribution of the channel, while the peak deviation amplifies local significant peak responses, enabling routing decisions to simultaneously perceive both global distribution and local saliency. This allows for precise dynamic operator selection for small low-contrast defects and defects of different morphologies. Top-2 sparse selection activates only a few expert branches, and combined with the low computational cost of the identity mapping expert in simple background regions, ensures that the module's computational load is controllable. Compared to traditional hybrid expert schemes that rely solely on a single statistic (such as the global average) for routing decisions, this embodiment uses a dual-statistical descriptor composed of the spatial mean and peak deviation to make routing decisions more accurate, avoiding feature mismatch problems caused by fuzzy routing and improving the representation ability of multi-morphological defects.

[0086] In conjunction with the dual statistical sparse hybrid expert module, the strip depthwise convolution module is specifically designed for slender defects with extremely high aspect ratios, and its structure is as follows: Figure 4 As shown. Input feature map Two parallel strip-shaped depthwise convolution branches are fed in simultaneously.

[0087] The horizontal branch uses a depthwise convolution with a kernel size of 1×k, and the number of groups is equal to the number of channels C of the input feature map, capturing horizontal strip textures along the width direction: The first processed feature is output after batch normalization and SiLU activation.

[0088] The vertical branch uses a depthwise convolution with a kernel size of k×1, and the number of groups is equal to the number of channels C of the input feature map, capturing vertical strip textures along the height direction: After batch normalization and SiLU activation, the second processed feature is output. The first processed feature and the second processed feature are added together, and a bidirectional strip feature is fused. Then, after depthwise separable convolution, the output is...

[0089] in, and Both are depthwise convolutions with the number of groups equal to the number of input channels C.

[0090] It should be noted that slender defects such as cracks and scratches are distributed in strips with a very large aspect ratio in the image. The receptive field shape of the standard square convolution kernel does not match the defect morphology and cannot obtain enough contextual information along the direction of defect extension, resulting in incomplete detection of slender defects and easy to miss detection. In contrast, the strip convolution kernel extends in a single direction, and its receptive field shape naturally matches the morphology of slender defects. The reason for using parallel horizontal and vertical branches is that the extension direction of cracks and scratches in the image is uncertain. Single-direction strip convolution can only cover defects in a single direction, while parallel dual branches can cover both horizontal and vertical directions at the same time. The two outputs are added together to achieve the fusion of bidirectional strip features. The reason for using depthwise convolution is that the number of groups in depthwise convolution is equal to the number of input channels, and each channel is convolved independently. The number of parameters is much smaller than that of standard convolution. When k is a large value (such as 7), the number of parameters of strip depthwise convolution is only 2k / 9≈1.56 times that of 3×3 depthwise convolution, which can obtain a long-distance strip receptive field with extremely low parameter overhead. When k is set to 7, the receptive field of strip convolution can cover the entire extension direction of the crack or scratch along a single direction, balancing coverage and parameter overhead. Each branch undergoes batch normalization and SiLU activation to stabilize the training process of each branch and introduce nonlinear expressive power. The addition of two branches achieves the fusion of bidirectional strip features. The terminal DWConv depthwise separable convolution is used to adjust the number of output channels to match the subsequent feature dimensions. Thus, the strip depthwise convolution module enhances the directional contextual expression of defects with large aspect ratios such as cracks and scratches with extremely low parameter overhead, effectively reducing the false negative rate of slender defects.

[0091] Through the collaborative work of the two modules mentioned above, this embodiment improves the feature representation capability of weak texture defects (such as small low-contrast defects) and slender defects (such as cracks and scratches), and realizes adaptive feature extraction and enhancement for defects of different shapes, thereby improving detection accuracy while maintaining the lightweight nature of the network.

[0092] On the other hand, such as Figure 5 As shown, this invention also provides an aircraft skin defect detection system based on dual statistical sparse experts and stripe depth convolution, including an image acquisition module, a defect detection module, and a detection report generation module. The image acquisition module is used to acquire aircraft skin images. It can employ a camera mounted on a drone, a camera built into a handheld inspection device, or a fixed industrial camera in a hangar. Drone inspections can cover the hangar exterior surface and high-altitude areas; handheld inspection devices are suitable for close-range manual re-inspection; and fixed hangar cameras are suitable for large-scale automatic acquisition at scheduled inspection stations. These three methods complement each other and can meet the needs of different inspection scenarios. The defect detection module uses the detection network trained in the above embodiments to perform defect detection on the acquired aircraft skin images, outputting the defect type and location in real time. Since the detection network uses end-to-end NMS-Free inference and has fewer than 3M parameters (i.e., 3*10), the system is highly efficient. 6The system features low inference latency, meeting the requirement for real-time defect detection during inspections. The inspection report generation module generates reports based on defect type and location. Examples of report content include defect category, location coordinates, confidence level, and skin part number. Its purpose is to output inspection results in a structured format, facilitating defect location and repair scheduling by maintenance personnel. The defect detection module can be deployed in three forms: UAV inspection system (airborne edge computing unit), handheld inspection instrument (mobile inference), and hangar fixed vision inspection system (industrial computer inference), all meeting real-time detection requirements. After acquiring skin images, the acquisition end loads detection network weights, outputs defect type and location in real time, and generates an inspection report.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting aircraft skin defects based on dual statistical sparse experts and stripe depth convolution, characterized in that, Includes the following steps: Obtain aircraft skin images; The aircraft skin image is input into a YOLO26-based detection network, which includes a backbone network, a neck network, and a detection head. A dual statistical sparse hybrid expert module is set in the backbone network and / or the neck network. The expert pool of the dual statistical sparse hybrid expert module includes at least two of the following: point defect experts, slender defect experts, fuzzy boundary experts, and identity mapping experts. The dual statistical sparse hybrid expert module calculates the spatial mean and spatial peak value of the input feature map, constructs a dual statistical descriptor based on the spatial mean and spatial peak value, inputs the dual statistical descriptor into a lightweight router to obtain the routing score of each expert, performs normalization processing and Top-k sparse selection on the routing score to obtain sparse expert weights, and sums the outputs of at least one selected expert branch in the expert pool according to the sparse expert weights to obtain the output feature map of the dual statistical sparse hybrid expert module. A strip deep convolution module is set in the high-resolution feature path of the neck network. The strip deep convolution module performs horizontal strip deep convolution and vertical strip deep convolution in parallel, and the output of the horizontal strip deep convolution is fused with the output of the vertical strip deep convolution to obtain an output feature map that enhances the directional contextual expression of slender defects. The detection network outputs the detection results of aircraft skin defects through the detection head, and the detection results include the defect type and defect location.

2. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The construction of the dual statistical descriptor based on the spatial mean and the spatial peak value includes: concatenating the spatial mean and the peak value deviation to construct the dual statistical descriptor, wherein the spatial mean... and spatial peak Calculate using the following formula: in, This represents the input feature map, where C, H, and W represent the number of channels, height, and width of the input feature map, respectively. These are the channel index, height index, and width index, respectively. The spatial mean and peak deviation are concatenated to construct the dual statistical descriptor. : The lightweight router is a two-layer multilayer perceptron (MLP), comprising a linear transformation, an activation function, and a linear output connected in sequence, used to nonlinearly map the dual statistical descriptor S to routing scores from each expert. , ,in, This represents the total number of experts in the expert pool.

3. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The process of normalizing the routing scores and obtaining sparse expert weights through Top-k sparse selection specifically includes: performing Softmax normalization on the routing scores to obtain normalized weights, selecting the top two experts with the largest normalized weights, resetting the weights of the unselected experts to zero, and renormalizing the non-zero weights to obtain sparse expert weights.

4. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The expert pool includes: Point defect expert, for 3×3 depthwise separable convolution; Slender Defect Expert, for strip-shaped depthwise convolutional modules; Fuzzy boundary expert, which is a dilated depthwise convolution with a dilation rate of 2, and connected to a 1×1 convolution; An expert in identity mapping, specializing in identity mapping.

5. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, In the dual statistical sparse hybrid expert module, the input feature map X is reduced in dimensionality by a shared 1×1 preprocessing convolution to obtain the intermediate feature Z. Each expert in the expert pool processes the intermediate feature Z. The outputs of the selected expert branches are weighted and summed according to the sparse expert weights, then increased in dimensionality by a 1×1 convolution, and finally added to the result obtained by shortcut connection of the input feature map X to serve as the output feature map of the dual statistical sparse hybrid expert module.

6. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The horizontal bar depthwise convolution is a depthwise convolution with a kernel size of 1×k, and the vertical bar depthwise convolution is a depthwise convolution with a kernel size of k×1. The number of groups in both the horizontal and vertical bar depthwise convolutions is equal to the number of channels C of the input feature map. The first processed feature obtained by batch normalization and SiLU activation of the horizontal bar depthwise convolution is added to the second processed feature obtained by batch normalization and SiLU activation of the vertical bar depthwise convolution, and then output after depthwise separable convolution.

7. The aircraft skin defect detection method based on dual statistical sparse experts and strip depth convolution as described in claim 1, characterized in that, The backbone network sequentially includes a first convolution, a second convolution, a first feature fusion module, a third convolution, a second feature fusion module, a fourth convolution, a third feature fusion module, a fifth convolution, an SPPF module, and a fourth feature fusion module; wherein, the first feature fusion module, the second feature fusion module, and the fourth feature fusion module are C2f modules, the third feature fusion module is a feature fusion module with 512 output channels, and the original C3k2 module or C2PSA module at the position of the third feature fusion module is replaced by the dual statistical sparse hybrid expert module; the SPPF module is an SPPF module with shortcut links.

8. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The neck network adopts an FPN-PAN structure, including an FPN upsampling path and a PAN downsampling path. The neck network outputs P3 features, P4 features and P5 features, where P3 features, P4 features and P5 features are the highest resolution features, medium resolution features and lowest resolution features, respectively, and correspond to the detection of small targets, medium targets and large targets, respectively. The high-resolution feature path is the highest resolution P3 generation path of the neck network, and the strip-shaped depth convolution module is set in the P3 generation path. At least one C2f module in the neck network is replaced with the dual statistical sparse hybrid expert module; wherein, the output feature map of the dual statistical sparse hybrid expert module is used as a feature map in the backbone network and / or the neck network to participate in feature extraction and feature fusion, and the output feature map of the strip deep convolution module is used as a feature map in the high-resolution feature path of the neck network to participate in feature fusion; The neck network outputs the processed feature map to the detection head; The detection head includes three detection branches, which perform defect detection on features P3, P4 and P5 respectively, and output the detection results of aircraft skin defects.

9. The aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in claim 1, characterized in that, The detection network is a YOLO26-nano model, and the detection network adopts an end-to-end NMS-Free dual-head architecture; the defect types include cracks, dents, missing rivet heads, paint peeling, and scratches.

10. An aircraft skin defect detection system based on dual statistical sparse experts and stripe depth convolution, characterized in that, include: The image acquisition module is used to acquire images of the aircraft skin. The defect detection module is used to perform defect detection on the aircraft skin image using the aircraft skin defect detection method based on dual statistical sparse experts and stripe depth convolution as described in any one of claims 1 to 9, and output the defect type and defect location. The inspection report generation module is used to generate an inspection report based on the defect type and defect location.

Images