Textile Defect Detection Based on Multi-proportion Spatial Pyramid Convolution and Adaptive Multi-scale Feature Fusion

This paper proposes a textile defect detection method based on YOLOv8, aiming to enhance the algorithm’s adaptability through the introduction of deformable convolution modules, a multi-proportion spatial pyramid convolutional structure, and a novel feature fusion strategy. Firstly, the deformable convolution modules replace the original convolution modules in the backbone network, enabling the network to dynamically adjust the convolution kernel size and enhance the capability to capture features of defects with varying sizes. Secondly, the multi-proportion spatial pyramid convolution structure extracts richer defective feature information by concatenating multiple 1 × 3 and 3 × 1 convolution kernels to enhance the model’s sensitivity to different proportion scale properties. Finally, the new feature fusion strategy integrates the ASFF structure, efficiently capturing defect information across multiple scales. The experimental results show that the algorithm in this paper effectively solves the challenges of unbalanced aspect ratio and difficulty in detecting small defects, and provides an efficient and robust solution for textile defect detection. The accuracy of the proposed method is improved by 8.6% on the ZJU-Leaper dataset and 12.9% on the Tianchi dataset compared with the baseline model.