改进YOLOv5网络的轻量级服装目标检测方法

为进一步降低基于深度学习的服装目标检测模型对计算资源的占用,提出一种改进的轻量级服装目标检测方法MV3L-YOLOv5。首先使用移动网络MobileNetV3_Large构造YOLOv5的主干网络;然后在训练阶段使用标签平滑策略,以增强模型泛化能力;最后使用数据增强技术弥补DeepFashion2数据集中不同服装类别图像数量不均衡问题。实验结果表明:MV3L-YOLOv5的模型体积为10.27 MB,浮点型计算量为10.2×10⁹次,平均精度均值为76.6%。与YOLOv5系列最轻量的YOLOv5s网络相比,模型体积压缩了26.4%,浮点型计算量减少了39%,同时平均精度均值提高了1.3%。改进后的算法在服装图像的目标检测方面效果有所提升,且模型更加轻量,适合部署在资源有限的设备中。

Lightweight clothing detection method based on an improved YOLOv5 network

In order to further reduce the occupation of computing resources by the clothing object detection model based on deep learning, an improved lightweight clothing object detection method, MV3L-YOLOv5, was proposed. The MobileNetV3_Large is used to construct the backbone network of YOLOv5, and the label smoothing strategy was introduced to enhance the generalization ability at the training stage of the model. The data augmentation technology was used to make up for the unbalanced number of images of different clothing categories in the DeepFashion2 dataset. Experimental results show that the model volume of MV3L-YOLOv5 is 10.27 MB, the floating-point operations is 10.2×10⁹ times, and mean average precision is 76.6 %. Comparing with YOLOv5s, which is the lightest network in YOLOv5 series, MV3L-YOLOv5 is compressed in the model volume by 26.4 %, reduced the floating-point operations by 39 %, and improved accuracy by 1.3 %. Experimental results in the improved algorithm show that the detection performance is notably improved, and the model is lighter and more suitable for deployment in devices with limited resources.