Abstract:

Objective Deep learning models have poor generalization performance when faced with cross-scene fabric defect detection, and there is relatively little research on dynamic cross scene transfer methods. Due to the influence of camera type, parameters, environmental lighting and other image acquisition conditions, there are significant differences in the distribution of fabric image data. How to accurately extract target domain data under various imaging conditions and achieve effective detection of fabric defects across scenes is an urgent problem to be solved. To this end, a hybrid algorithm based on the Beetle global search algorithm(BAS) was constructed by adding local search capabilities to the global search capability of BAS to tackle the complexity and diversity in fabric images during fabric defect detection.

Method This research constructed a hybrid algorithm based on the Beetle algorithm by adding local search capabilities to the global search capability of BAS. Gabor filters were used to select the optimal parameters and establish a fabric detection scheme. In order to solve the optimization problem of the Beetle algorithm, local search capability was added to the global search capability of BAS. In order to obtain accurate binarization detection results, the image underwent threshold segmentation based on the use of low-pass filtering to convolution the results again.

Results In order to verify the effectiveness of the BAS model, the method proposed in this paper was compared with the methods in references. It was seen that the accuracy curve of this method was improved fastest and took the shortest time to reach the maximum accuracy, but the loss function curve fluctuated greatly. To verify the accuracy of the proposed method, the T-SNE method was used to visualize the features of fabric defects using the improved BAS method and the methods used in references. The method in this article showed a smaller distance in the embedding space, but the feature similarity extracted from defects and hole defects at the edge of the image was higher. To verify whether the proposed method extracted the features of defects, the Grad CAM method was used to visualize the defect features extracted by the model. The focus of this method was more concentrated within the range of fabric defects, and it was less affected by the background area. This verified that this method could effectively identify defect areas and has good generalization performance for cross-scene fabric defect detection. The optimal Gabor filter was used for fabric defect detection, and the defect detection effect of the proposed algorithm was evaluated in the form of binary images. The results show that the fabric defects detected by this algorithm are clear and accurate.

Conclusion The loss function curve of the method in this paper converge quickly, and the accuracy curve convergence value is high. The space for extracting all defect features of defects is closer, and the distance between each defect is larger. This proves that the features extracted from fabric defects in this paper are classified differently, and are less affected by the background. It verifies that the method in this paper can effectively identify the defect area with good generalization performance for cross-scene fabric defect detection. In practical applications, obtaining fabric images is influenced by lighting and fabric texture, which can affect the applicability of the model across-scenes, such as camera angle, fabric type, and camera parameters. The next step is to conduct research on other influencing factors such as camera angle and fabric type to further improve the model's generalization ability across scenes.

Key words: deep learning, global search algorithm, Gabor filter, fabric defect detection, generalization performance, pattern recognition