An aircraft skin surface anomaly detection method and system independent of defect samples and a storage medium

Abstract

This invention discloses a method, system, and storage medium for anomaly detection on aircraft skin surfaces that does not rely on defect samples. Belonging to the technical field of visual inspection in aerospace manufacturing, the method involves acquiring images of the aircraft skin surface and dividing them into several image blocks. Based on a multimodal large model, it obtains the semantic feature vectors, normal text feature vectors, and anomalous text feature vectors of the image blocks. The similarity between the semantic feature vectors and the text feature vectors is calculated to evaluate the semantic anomaly score. The deviation of the texture features of the image blocks from their normal distribution is analyzed to evaluate the texture anomaly score. Based on the semantic anomaly score and the texture anomaly score, a weighted fusion is performed to obtain a comprehensive anomaly score. If the comprehensive anomaly score is greater than or equal to a set threshold, a pixel-level anomaly heatmap is generated based on the attention map output by the visual encoder of the multimodal large model, and the anomaly region is located. This invention achieves interpretable and localizable anomaly detection on the skin surface.

Description

Technical Field

[0001] This invention belongs to the technical field of visual inspection in aerospace manufacturing, specifically relating to a method, system, and storage medium for detecting anomalies on aircraft skin surfaces that do not rely on defect samples. Background Technology

[0002] In the field of aerospace manufacturing, aircraft skin surfaces are critical load-bearing and aerodynamic components; surface abnormalities can affect structural safety, fatigue life, and aerodynamic efficiency. Therefore, surface anomaly detection is necessary throughout aircraft manufacturing and maintenance. Traditional aircraft skin surface quality inspection typically relies on manual visual inspection. Inspectors must assess the presence of scratches, dents, or other abnormalities under conditions of strong light reflection, complex curvature, and different coating materials. However, manual visual inspection is not only heavily influenced by subjective experience but also prone to missed or false detections in large areas of skin, making it difficult to guarantee consistency and traceability.

[0003] Currently, with the development of intelligent vision technology, supervised learning-based target detection methods have been introduced into the scenario of skin surface detection. For example, the existing Chinese patent CN102928435A, "A Method and Apparatus for Recognizing Aircraft Skin Damage Based on Image and Ultrasonic Information Fusion," pre-acquires images and ultrasonic echo signals of aircraft skin for each known damage category, extracts the texture features of the images and the ultrasonic echo features, and then trains a classifier using the feature vector composed of the image texture features and ultrasonic echo features as input and the damage category as output. Similarly, images and ultrasonic echo signals of the aircraft skin to be identified are acquired, and the image texture features and ultrasonic echo features are extracted. The feature vector composed of the image texture features and ultrasonic echo features is then input into the trained classifier, and the output of the classifier is the damage category of the aircraft skin to be identified.

[0004] However, such methods rely on a large amount of defect-annotated data for training, while defect samples are scarce and defect morphologies are diverse in aerospace manufacturing scenarios, resulting in insufficient generalization ability of supervised learning-based detection models. Secondly, supervised learning-based target detection models mainly rely on explicit appearance features, which are sensitive to changes in lighting, reflective contamination, and minute surface texture anomalies, making it difficult to adapt to the complex environment of actual field operations. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and storage medium for detecting anomalies on aircraft skin surfaces that do not rely on defect samples, in order to solve the aforementioned problems.

[0006] This invention is mainly achieved through the following technical solutions: A method for detecting anomalies on aircraft skin surfaces that does not rely on defect samples, comprising the following steps: Step S1: Image acquisition and processing; acquire images of the aircraft skin surface and divide the images into several image blocks; Step S2: Based on the multimodal large model, obtain the image semantic feature vector, normal text feature vector, and abnormal text feature vector of the image patch. Calculate the similarity between the image semantic feature vector and the normal text feature vector and the abnormal text feature vector, respectively, and evaluate the semantic abnormality score based on the difference between the two similarities. Step S3: Extract the texture features of the image patch and characterize the degree of deviation of the texture features from the normal distribution by Mahalanobis distance, and then evaluate the texture anomaly score; Step S4: Based on the semantic anomaly score in Step S2 and the texture anomaly score in Step S3, perform weighted fusion to obtain the comprehensive anomaly score of the image patch; if the comprehensive anomaly score is greater than or equal to a set threshold, mark the image patch as a potential defect area and proceed to Step S5; otherwise, mark the image patch as a normal area. Step S5: For potential defect areas, generate pixel-level anomaly heatmaps based on the attention map output by the visual encoder of the multimodal large model, which are used to accurately locate the anomaly areas. Step S6: Perform threshold segmentation on the abnormal heat map to obtain the binary mask of the abnormal region, obtain the location information of the abnormal region based on the binary mask of the abnormal region, and output the defect location information.

[0007] Preferably, the comprehensive anomaly score in step S4 is used to evaluate the degree of anomaly of the image patch on the aircraft skin surface. If the comprehensive anomaly score is greater than or equal to a set threshold, step S5 is executed; otherwise, the image patch is marked as a normal area. Specifically, step S4 is a prerequisite for step S5. Step S4 aims to filter out defect areas with high comprehensive anomaly scores, and then step S5 is performed to generate pixel-level anomaly heatmaps for the areas with high comprehensive anomaly scores to locate the anomaly areas. Finally, in step S6, the location information of the defects and the comprehensive anomaly score are output.

[0008] To better realize the present invention, the multimodal large model further includes a visual encoder and a text encoder, and step S2 includes the following steps: Step S21: Input the image patch into the visual encoder and output the image semantic feature vector; Step S22: Input the normal skin surface cue words and the abnormal skin surface cue words into the text encoder respectively, and output the normal text feature vector and the abnormal text feature vector accordingly; Step S23: Calculate the semantic similarity between the image semantic feature vector and the normal text feature vector, and between the image semantic feature vector and the abnormal text feature vector, respectively, and evaluate the semantic anomaly score based on the difference between the two similarities.

[0009] To better implement the present invention, further, in step S23, the normal semantic similarity between the image semantic feature vector and the normal text feature vector is calculated based on cosine similarity. s ( i , n (and the abnormal semantic similarity between image semantic feature vectors and abnormal text feature vectors), and the abnormal semantic similarity between them. s ( i , a ); ; ; in: It is the semantic feature vector of the image; t n These are normal text feature vectors; t a This is the feature vector of the abnormal text.

[0010] To better implement the present invention, further, in step S23, the semantic anomaly scoring... for: ; in: s 1 is the monotonic normalization function.

[0011] To better realize the present invention, step S3 further includes the following steps: Step S31: Construct a training set based on normal skin samples, and construct a feature distribution model of normal texture features based on the training set; Based on a self-supervised texture encoder, the normal feature set of the training set is obtained { z 1, z 2,…, z M Then, based on the normal feature set, a feature distribution model of normal texture features is constructed. in: z M Let M be the texture feature of the Mth normal skin sample, where M is the total number of normal skin samples; Step S32: Extract texture features of image patches based on a self-supervised texture encoder; then, based on the feature distribution model, characterize the degree of deviation of the texture features of the image patches from the normal distribution using Mahalanobis distance; Step S33: Evaluate the texture anomaly score based on the degree of deviation.

[0012] To better implement this invention, in step S31, the feature distribution model of normal texture features is further defined as follows: ; ; in: m This is the mean vector of the normal feature set; ∑ is the covariance matrix of the normal feature set; z j For the first j Texture features of a normal skinned sample; T is the matrix transpose.

[0013] To better implement the present invention, further, in step S32, the texture feature z of the image patch... i Degree of deviation from the normal distribution for: ; in: m This is the mean vector of the normal feature set; ∑ is the covariance matrix of the normal feature set; z i Texture features of image patches; In step S33, the texture anomaly score is: ; in: Scoring for texture anomalies; s 2 is the distribution normalization function.

[0014] To better implement the present invention, in step S5, the multi-layer self-attention weights output by the visual encoder of the multimodal large model are used to obtain the pixel-level attention map of the image block through the Grad-CAM method. Then, the pixel-level attention map is normalized to obtain the pixel-level anomaly heatmap.

[0015] To better implement the present invention, further, in step S6, a binary mask is generated based on the pixel-level anomaly heatmap using an adaptive thresholding method; then, the location and bounding box of the anomaly region are extracted; finally, the mask, bounding box, and comprehensive anomaly score of the anomaly region are output.

[0016] An aircraft skin surface anomaly detection system independent of defect samples, implemented based on the aforementioned aircraft skin surface anomaly detection method independent of defect samples, includes: The image acquisition and processing module is used to acquire images of the aircraft skin surface and divide them into several image blocks; The semantic anomaly scoring and evaluation module is used to obtain the image semantic feature vector, normal text feature vector and abnormal text feature vector of the image patch based on the multimodal large model, and evaluate the semantic anomaly score by the similarity between the semantic feature vector and the text feature vector. The texture anomaly scoring module is used to characterize the degree of deviation of texture features from the normal distribution using Mahalanobis distance, and to evaluate the texture anomaly score based on the degree of deviation. The anomaly score fusion module is used to weight and fuse semantic anomaly scores and texture anomaly scores to obtain a comprehensive anomaly score; The heatmap generation module is used to generate pixel-level anomaly heatmaps from the attention map output by the visual encoder based on a multimodal large model. The defect output module is used to generate a mask of the abnormal region based on the abnormal heat map, obtain the location information of the abnormal region based on the mask, and finally output the defect location information.

[0017] Preferably, the system can be deployed in actual production environments using edge deployment, cloud deployment, or hybrid deployment methods, and integrated with systems such as MES to achieve defect recording and traceability. Edge deployment uses embedded devices to process image blocks for real-time on-site detection; cloud deployment batches images and uploads them to a cloud server for unified inference, enabling large-scale skin inspection tasks in the hangar; hybrid deployment completes pre-inspection through inspection terminals, while detailed analysis and quality archiving are performed in the cloud.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting anomalies on aircraft skin surfaces that is independent of defect samples.

[0019] The beneficial effects of this invention are as follows: (1) This invention does not require defect samples. Through multi-source fusion of semantic and texture deviations, it achieves robust identification of abnormal regions, effectively improving the consistency and reliability of aircraft skin surface anomaly detection in complex detection scenarios. This invention provides pixel-level anomaly localization results through anomaly heatmaps, and can output anomaly region masks, bounding boxes, and heatmaps, which can be used for manual verification and interface with quality traceability systems, meeting the strict requirements of the aviation industry for interpretability and traceability. This invention solves the problems of sample scarcity, light sensitivity, weak generalization ability, and insufficient interpretability in existing aircraft skin surface anomaly detection, and realizes interpretable and localizable anomaly detection on the skin surface.

[0020] (2) This invention can establish a feature distribution model based solely on normal skin samples, without the need to collect a large number of defect samples or to perform detailed annotation of defects. This invention adopts a self-supervised method from semantic feature and texture feature extraction to normal distribution modeling, avoiding the dependence on manual annotation, and is especially suitable for aerospace manufacturing scenarios where defect samples are extremely scarce.

[0021] (3) This invention achieves reliable identification of surface anomalies on aircraft skin by weighted fusion of semantic anomaly scores and texture anomaly scores. It has good generalization ability to aircraft skin of different materials and can maintain stable detection performance under different lighting, reflection and shadow conditions. Specifically, this invention adaptively integrates semantic anomaly scores and texture anomaly scores based on a weighted fusion mechanism, which improves robustness to complex lighting and false defects and enhances the generalization ability of detection.

[0022] (4) This invention realizes semantic-texture dual-dimensional interpretability analysis. It calculates the semantic similarity difference between image blocks and normal / abnormal prompt words through a multimodal large model to evaluate the semantic level of abnormality score. Based on the self-supervised texture encoder, it constructs a feature distribution model of normal texture, quantifies the local texture deviation, and provides a physically interpretable basis for texture anomalies. Attached Figure Description

[0023] Figure 1 This is a flowchart of the aircraft skin surface anomaly detection method of the present invention that does not rely on defect samples; Figure 2 Example images for detecting anomalies on aircraft skin surfaces, along with pixel-level anomaly heatmaps of the anomalous areas; Among them: (a) is an example of an image of an anomaly on the surface of the aircraft skin; (b) Example 2 of an image showing anomalies on the surface of an aircraft skin; (c) is an example three of images showing anomalies on the surface of aircraft skin; (d) is the pixel-level anomaly heatmap corresponding to image example one in (a); (e) is the pixel-level anomaly heatmap corresponding to image example two in (b); (f) is the pixel-level anomaly heatmap corresponding to image example three in (c). Detailed Implementation

[0024] Example 1: A method for detecting anomalies on aircraft skin surfaces that does not rely on defect samples, such as Figure 1 As shown, it includes the following steps: Step S1: Image acquisition and processing to obtain several image blocks; acquire an image of the target aircraft skin surface, and divide the image into multiple image blocks according to a preset sliding window size and overlap rate for subsequent local feature analysis.

[0025] Specifically, an image acquisition system is used to acquire images of the target aircraft skin surface. The image acquisition system mainly includes a high-definition industrial camera, a portable inspection terminal or robot, and a ring light source. The ring light source has a zone function, and the industrial camera has automatic exposure and automatic gain functions.

[0026] During image acquisition, the camera is kept at a top-down angle to be as perpendicular as possible to the skin surface, and the light source brightness is adjusted to avoid areas of high-saturation highlights. After image acquisition, the acquired images are preprocessed using brightness normalization, histogram equalization, or normalization transformations. After image preprocessing, the acquired high-resolution image is divided into sections using a sliding window method. N Image patches. For example, the image patch size can be set to 512 pixels, and the sliding window overlap rate can be set to 40% to ensure that minor or local defects are not missed due to the slice boundaries.

[0027] Step S2: Evaluate the semantic anomaly score; Based on the multimodal large model, obtain the image semantic feature vector, normal text feature vector, and abnormal text feature vector of the image patch, calculate the similarity between the image semantic feature vector and the normal text feature vector and the abnormal text feature vector respectively, and evaluate the semantic anomaly score based on the difference between the two similarities.

[0028] Preferably, step S2 includes the following steps: (1) Input the image patch into the visual encoder of the multimodal large model to obtain the corresponding image semantic feature vector. Specifically, the image patch x i Visual encoders using multimodal large models E vlm Obtain the semantic feature vector of the image ,in i ∈1,2,3,…, N ; (2) Input the normal skin surface cue words and the abnormal skin surface cue words into the text encoder of the multimodal large model respectively to obtain the normal text feature vector and the abnormal text feature vector. Specifically, the normal skin surface cue words... p n and abnormal skin surface warning words p a Text encoder E text Normal text feature vectors were obtained respectively. and abnormal text feature vectors .

[0029] (3) Calculate the semantic feature vectors of the images respectively. Compared with normal text feature vectorst n and image semantic feature vectors With abnormal text feature vectors t a The cosine similarity corresponds to the normal semantic similarity. and abnormal semantic similarity .

[0030] (4) Based on normal semantic similarity s ( i , n ) and abnormal semantic similarity s ( i , a The difference between the two values ​​is used to evaluate the semantic anomaly score. ,in s 1 represents the monotonic normalization function. It can be used to assess whether image patches deviate from the normal skinning state at the semantic level based on semantic anomaly scoring.

[0031] Preferably, the multimodal large model visual encoder E vlm and text encoder E text Specifically, this refers to the visual encoder and text encoder of the CLIP model. Specifically, after inputting an image patch, it outputs an image semantic feature vector with a dimension of 1024; after inputting a prompt word, it outputs a text feature vector with a dimension of 1024.

[0032] Preferably, the description of a normal skin surface includes terms such as "normal skin surface", "undamaged surface", and "complete coating"; the description of an abnormal skin surface includes terms such as "surface scratches", "uneven coating", "dents", "corrosion", and "foreign matter on the surface".

[0033] Preferably, considering the diversity of aircraft skin materials, dynamic material cue word injection is designed. Specifically, before inputting the text encoder, the material attributes are dynamically injected into cue words based on the current detected skin type, such as aluminum alloy, composite materials, etc. For example, the general cue word "normal skin surface" can be dynamically adjusted to "normal skin surface with carbon fiber weave texture" or "smooth aluminum alloy primer surface". By refining the text description, this invention enables the text feature vector to more accurately match the physical characteristics of the current skin, reducing semantic misjudgments caused by differences in material texture.

[0034] Step S3: Evaluate the texture anomaly score; Extract texture features of image patches based on a self-supervised texture encoder, and characterize the degree of deviation of texture features from the normal distribution by Mahalanobis distance, thereby evaluating the texture anomaly score.

[0035] Preferably, step S3 includes the following steps: (1) Extract features from image patches using a self-supervised texture encoder to obtain texture feature representations; for example, image patches x i Through texture encoder E text Obtain texture features ; (2) Establish a feature distribution model of normal texture features on a training set containing only normal skin samples. Specifically, obtain the normal feature set of the training set { through a self-supervised texture encoder. z 1, z 2,…, z M}; Then, using the normal feature set { z 1, z 2,…, z M} Calculate the mean vector μ and covariance matrix ∑ to obtain the feature distribution model of normal texture features.

[0036] ; ; in, z j For the first j Texture features of a normal skinned sample.

[0037] Preferably, images of normal conditions are collected from different skin surfaces as normal skin samples.

[0038] Preferably, a self-supervised texture encoder is trained using an unlabeled method. The self-supervised texture encoder is obtained by randomly occluding and reconstructing the input image based on the MAE model.

[0039] (3) Finally, the deviation of the texture features of the image patch from the normal distribution is calculated to obtain the texture anomaly score.

[0040] 1) Representing texture features using Mahalanobis distance z i The degree of deviation from normal texture features .

[0041] 2) Based on the degree of deviation Evaluation of texture anomaly score ,in s 2 is the distribution normalization function.

[0042] Step S4: Based on the semantic anomaly score in Step S2 and the texture anomaly score in Step S3, a weighted fusion is performed to obtain a comprehensive anomaly score for the image patch. Specifically, the semantic anomaly score and the texture anomaly score are weighted and fused according to preset weights to obtain a comprehensive anomaly score, achieving joint discrimination from multiple information sources. The higher the comprehensive anomaly score, the greater the anomaly probability. If the comprehensive anomaly score is ≥ a set threshold, the image patch is marked as a potential defect region, and the process proceeds to Step S5; otherwise, the image patch is marked as a normal region.

[0043] Preferably, based on semantic anomaly scoring and texture anomaly scoring The overall anomaly score is obtained as follows: ; in, w 1 and w 2 represents the scoring weights, and w 1+ w 2 = 1.

[0044] Step S5: For potential defect areas, generate pixel-level anomaly heatmaps based on the attention map output by the visual encoder of the multimodal large model, which are used to finely locate the anomaly areas. Specifically, based on the multi-layer self-attention weights of a multimodal large-scale visual encoder, pixel-level attention maps of image patches are obtained through the Grad-CAM method. After normalization, pixel-level attention maps yield pixel-level anomaly heatmaps, which are used to locate anomalous regions. Anomaly regions can be extracted based on these pixel-level anomaly heatmaps to characterize potentially anomalous areas in the image.

[0045] Step S6: Threshold segmentation is performed on the anomaly heatmap to obtain a binary mask of the anomaly region, thereby acquiring the location information of the anomaly region and outputting defect detection information. The defect detection information includes defect location information and a comprehensive anomaly score.

[0046] Specifically, a binary mask is generated from the pixel-level anomaly heatmap using an adaptive thresholding method; and the location and bounding box of the anomaly region are extracted through morphological operations such as connected component analysis and erosion / dilation. The output includes the mask, bounding box of the anomaly region, and a comprehensive anomaly score, which are used for recording and tracing in the quality inspection system.

[0047] like Figure 2 As shown, based on the specimen images of three types of aircraft skin surfaces (a), (b), and (c), the present invention generates pixel-level anomaly thermal maps (d), (e), and (f) respectively. The abnormal areas of the skin surface can be directly detected on the pixel-level anomaly thermal maps.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting anomalies on aircraft skin surfaces that does not rely on defect samples, characterized in that, Includes the following steps: Step S1: Image acquisition and processing; acquire images of the aircraft skin surface and divide the images into several image blocks; Step S2: Based on the multimodal large model, obtain the image semantic feature vector, normal text feature vector, and abnormal text feature vector of the image patch. Calculate the similarity between the image semantic feature vector and the normal text feature vector and the abnormal text feature vector, respectively, and evaluate the semantic abnormality score based on the difference between the two similarities. Step S3: Extract the texture features of the image patch and characterize the degree of deviation of the texture features from the normal distribution by Mahalanobis distance, and then evaluate the texture anomaly score; Step S4: Based on the semantic anomaly score in Step S2 and the texture anomaly score in Step S3, perform weighted fusion to obtain the comprehensive anomaly score of the image patch; if the comprehensive anomaly score is greater than or equal to a set threshold, mark the image patch as a potential defect area and proceed to Step S5; otherwise, mark the image patch as a normal area. Step S5: For potential defect areas, generate pixel-level anomaly heatmaps based on the attention map output by the visual encoder of the multimodal large model. Step S6: Perform threshold segmentation on the abnormal heat map to obtain the binary mask of the abnormal region, obtain the location information of the abnormal region based on the binary mask of the abnormal region, and output the defect location information.

2. The method for detecting surface anomalies on aircraft skin that does not rely on defect samples, as described in claim 1, is characterized in that, The multimodal large model includes a visual encoder and a text encoder, and step S2 includes the following steps: Step S21: Input the image patch into the visual encoder and output the image semantic feature vector; Step S22: Input the normal skin surface cue words and the abnormal skin surface cue words into the text encoder respectively, and output the normal text feature vector and the abnormal text feature vector accordingly; Step S23: Calculate the semantic similarity between the image semantic feature vector and the normal text feature vector, and between the image semantic feature vector and the abnormal text feature vector, respectively, and evaluate the semantic anomaly score based on the difference between the two similarities.

3. The method for detecting surface anomalies of aircraft skin that does not rely on defect samples, as described in claim 2, is characterized in that... In step S23, the normal semantic similarity between the image semantic feature vector and the normal text feature vector is calculated based on cosine similarity. s ( i , n (and the abnormal semantic similarity between image semantic feature vectors and abnormal text feature vectors), and the abnormal semantic similarity between them. s ( i , a ); ； ； in: It is the semantic feature vector of the image; t n These are normal text feature vectors; t a This is the feature vector of the abnormal text.

4. The method for detecting anomalies on aircraft skin surfaces that does not rely on defect samples, as described in claim 3, is characterized in that... In step S23, the semantic anomaly scoring for: ； in: σ 1 is the monotonic normalization function.

5. The method for detecting surface anomalies on aircraft skin that does not rely on defect samples, as described in claim 1, is characterized in that... Step S3 includes the following steps: Step S31: Construct a training set based on normal skin samples, and construct a feature distribution model of normal texture features based on the training set; Based on a self-supervised texture encoder, the normal feature set of the training set is obtained { z 1, z 2,…, z M Then, based on the normal feature set, a feature distribution model of normal texture features is constructed. in: z M Let M be the texture feature of the Mth normal skin sample, where M is the total number of normal skin samples; Step S32: Extract texture features of image patches based on a self-supervised texture encoder; then, based on the feature distribution model, characterize the degree of deviation of the texture features of the image patches from the normal distribution using Mahalanobis distance; Step S33: Evaluate the texture anomaly score based on the degree of deviation.

6. The method for detecting anomalies on aircraft skin surfaces independent of defect samples according to claim 5, characterized in that, In step S31, the feature distribution model of normal texture features is as follows: ； ； in: μ This is the mean vector of the normal feature set; ∑ is the covariance matrix of the normal feature set; z j For the first j Texture features of a normal skinned sample; T is the matrix transpose.

7. A method for detecting anomalies on aircraft skin surfaces that does not rely on defect samples, as described in claim 5 or 6, characterized in that, In step S32, the texture features z of the image patch i Degree of deviation from the normal distribution for: ； in: μ This is the mean vector of the normal feature set; ∑ is the covariance matrix of the normal feature set; z i Texture features of image patches; In step S33, the texture anomaly score is: ； in: Scoring for texture anomalies; σ 2 is the distribution normalization function.

8. The method for detecting surface anomalies in aircraft skin that does not rely on defect samples, as described in claim 1, is characterized in that, In step S5, the multi-layer self-attention weights output by the visual encoder based on the multimodal large model are used to obtain the pixel-level attention map of the image patch through the Grad-CAM method. Then, the pixel-level attention map is normalized to obtain the pixel-level anomaly heatmap.

9. An aircraft skin surface anomaly detection system independent of defect samples, implemented based on the aircraft skin surface anomaly detection method independent of defect samples as described in any one of claims 1 to 8, characterized in that, include: The image acquisition and processing module is used to acquire images of the aircraft skin surface and divide them into several image blocks; The semantic anomaly scoring and evaluation module is used to obtain the image semantic feature vector, normal text feature vector and abnormal text feature vector of the image patch based on the multimodal large model, and evaluate the semantic anomaly score by the similarity between the semantic feature vector and the text feature vector. The texture anomaly scoring module is used to characterize the degree of deviation of texture features from the normal distribution using Mahalanobis distance, and to evaluate the texture anomaly score based on the degree of deviation. The anomaly score fusion module is used to weight and fuse semantic anomaly scores and texture anomaly scores to obtain a comprehensive anomaly score; The heatmap generation module is used to generate pixel-level anomaly heatmaps from the attention map output by the visual encoder based on a multimodal large model. The defect output module is used to generate a mask of the abnormal region based on the abnormal heat map, obtain the location information of the abnormal region based on the mask, and finally output the defect location information.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the aircraft skin surface anomaly detection method according to any one of claims 1 to 8, which is independent of defect samples.

Images