基于改进YOLOv5的服装logo识别

THE RECOGNITION OF CLOTHING LOGO BASED ON ENHANCED YOLOV5

摘要: 使用改进的YOLOv5算法应用于常见的服装品牌logo识别，采用加权k-means算法聚类得出9个初始锚框，针对原始YOLOv5的主干网络，增加Transformer自注意力机制和SPPF快速空间金字塔池化。在现有的FlickrSportLogos-10数据集的基础上略作调整，删除低质量图片和相似图片，增加高质量图片进行实验。实验结果表明，改进后的YOLOv5算法的mAP@0.5和mAP@0.5:0.95分别达到90.7%和60.1%，相比原始YOLOv5分别提高了0.033和0.016。

Abstract: In this paper, the enhanced YOLOv5 algorithm is applied to clothing brand logo recognition. We used the weighted k-means algorithm to obtain 9 initial anchor boxes, and transformer self-attention mechanism and SPPF model were added to the original YOLOv5 backbone. We made some adjustments on the base of the existing FlickrSportlogos-10 dataset, such as removing some low quality and similar images and adding some high-quality images. The experimental results show that the mAP@0.5 and mAP@0.5:0.95 of the enhanced YOLOv5 algorithm reach 90.7% and 60.1%, which are 0.033 and 0.016 higher than the original YOLOv5 algorithm respectively.