Leather Defect Detection Based on Improved YOLOv8 Model

Abstract

Addressing the low accuracy and slow detection speed experienced by algorithms based on deep learning for a leather defect detection task, a lightweight and improved leather defect detection algorithm, dubbed YOLOv8-AGE, has been proposed based on YOLOv8n. In the backbone network, the EMA attention mechanism and C2f module have been fused into the C2f-EMA module, achieving performance enhancement with lower computational overhead. In the neck, the AFPN structure has been combined with the VoV-GSCSP module constructed using GSConv, to reduce the number of parameters while enhancing the model’s multi-scale detection capability. Finally, a shared convolutional layer has been introduced for simplifying the design of the detection head. Experimental results demonstrate that the improved algorithm achieves an improvement of 1.39% in mAP50 and reduces the number of parameters and GFLOPs by 9.3% and 7.41%, respectively, as compared to the original YOLOv8 model. On the dataset in this paper, there is an improvement in accuracy and detection speed over mainstream algorithms.