Yolov7-Tinier: Towards High-Precision and Lightweight Detection of Fabric Defects in Textile Plant

Abstract

To address the low recognition accuracy and poor real-time performance of models in online fabric defect detection tasks, an efficient and compact fabric defect detection method, YOLOv7-tinier, is introduced in this paper. YOLOv7-tinier makes several key improvements to the YOLOv7-tiny model. First, it uses partial convolution to reconstruct the feature extraction module ELAN in the backbone network, reducing the number of parameters and extracting more diverse and hierarchical features and thus improving the detection accuracy and speed. Secondly, a new module called Dilated Spatial Pyramid Pooling Fast Cross Stage Partial Concat is proposed to replace the original Spatial Pyramid Pooling Cross Stage Partial Concat, further reducing the number of parameters and improving the computational efficiency. Finally, it introduces a convolution structure with attention mechanism SConv(Self-attentional convolution) to replace the ordinary convolution of the Neck part, and SBL and ELAN-S modules are constructed, which substantially enhances the network’s detection accuracy without significantly increasing the number of parameters. Extensive comparison and ablation experiments were conducted on the real fabric defect dataset. The experimental results show that under identical conditions, YOLOv7-tinier, our proposed model, achieved a 9.55% improvement in mean Average Precision (mAP) and a 10.81% reduction in parameters compared to the baseline YOLOv7 model, while maintaining a Frames Per Second (FPS) rate of 155.27 Hz. This model can meet both the accuracy and real-time requirements of fabric defect detection in textile manufacturing environments.