基于YOLOv3的轻量化口罩检测算法研究

Research on Lightweight Mask Detection Algorithm Based on YOLOv3

针对当前基于深度学习的口罩检测算法在实时性与检测精度上不能同时具有良好的表现性能,本文提出一种基于YOLOv3的轻量化口罩检测算法,通过EfficientNet-B1网络替换掉原有的网络参数量大,网络结构复杂的骨干网络Darknet-53,为进一步提升网络性能,实验引入ECA通道注意力机制与特征金字塔结构相结合,最后采用CIOU对原有的边界框损失进行优化。实验结果表明,该网络结构模型与YOLOv3相比,检测精度仅降低1.73%,但模型参数量降低了79%,且单张图片检测速度也提升了3.93倍,一定程度上体现了本文算法的良好性能。

In view of the fact that the current deep learning-based mask detection algorithm cannot have good performance in real-time and detection accuracy at the same time, this thesis proposes a light-weight mask detection algorithm based on YOLOv3, which replaces the original backbone network Darknet-53 which has a large number of network parameters and a complex network structure through the EfficientNet-B1 network. In order to further improve the network performance, the experiment introduces the ECA channel attention mechanism combined with the feature pyramid structure, and finally uses CIOU to optimize the original bounding box loss. The experimental results show that the network structure model is compared with YOLOv3. The detection accuracy is only reduced by 1.73%, but the amount of model parameters is reduced by 79%, and the detection speed of a single image is also increased by 3.93 times, which reflects the good performance of the algorithm in this paper to a certain extent.