基于YOLOv4-GS检测算法的服装识别方法

Clothing Recognition Method Based on YOLOv4-GS Detection Algorithm

针对服装检测中检测模型体积大、运算量高,难以适应后续嵌入式平台环境的需要,本文提出了一种高效的YOLOv4-GS算法,极大降低了检测模型的规模。首先对数据集使用K-means聚类方法获得初始候选框,再融合Ghost模块和SimAM注意力机制组成GS模块,然后利用GS模块重构YOLOv4网络得到更轻量、更高效的YOLOv4-GS模型。实验结果表明:对比原生YOLOv4网络,在DeepFashion2数据集和相同环境下,YOLOv4-GS模型骨干网络浮点运算量减少48.33%,参数量减少49.63%,模型大小降低了33.12%,mAP达到67.8%,提升了2.1%。

Aiming at the needs that the detection model in clothing detection has large volume and high amount of computa⁃tion,which is difficult to adapt to the subsequent embedded platform environment,this paper proposes an efficient YOLOv4-GS algorithm,which greatly reduces the scale of the detection model.Firstly,the K-means clustering method is used to obtain the ini⁃tial candidate box for the datasets,and then the Ghost module and simAM attention mechanism are combined to form the GS mod⁃ule.Then,the GS module is used to reconstruct the YOLOv4 network to obtain a lighter and more efficient YOLOv4 GS model.The experimental results show that compared with the original YOLOv4 network,in the DeepFashion2 datasets and the same environ⁃ment,the YOLOv4-GS model backbone network FLOPs are reduced by 48.33%,the backbone network parameters are reduced by 49.63%,the model size is reduced by 33.12%,and mAP is reached 67.8%,an increase of 2.1%.