FM-YOLO Object Detection Algorithm
Xiaozhao Li, Nan Xiang, Chuanzhong Pan, Ruixian Zhang · 2022
Aiming at the problems of YOLOv4 with many parameters and low detection accuracy, a lightweight FM-YOLO (Fused Mobile-You Only Look Once) object detection algorithm is proposed. The algorithm improves the convolution layer of the deep network for feature extraction in two aspects. First, MBConv is used to expand the width of the convolution operation and reduce the amount of model parameters. Secondly, Fused-MBConv in the shallow network solves the problem that the depthwise separable convolution cannot fully utilize GPU acceleration in the shallow network, resulting in slower training speed. The activation function is changed from Mish to SiLU, and dropout is added to avoid network training over-fitting. From the test results of the VOC2007 data set, the parameter of the YOLOv4 algorithm is reduced by 1.4M compared with YOLOv4, and the mAP is increased by 1.45%. This algorithm improves the accuracy of detection while slightly reducing the amount of model parameters.