Ghost-YOLOX: A Lightweight and Efficient Implementation of Object Detection Model
Chunzhi Wang, Xin Tong, Jiahui Zhu, Rong Gao · 2022 26th International Conference on Pattern Recognition (ICPR) · 2022
In order to solve the problems of large number of parameters and high computational complexity in current object detection models, we propose a lightweight object detection model based on YOLOX. First, we use a lightweight network GhostNet as the backbone feature extraction network, and adjust the number of channels in the output feature layers of the backbone network using depth-separable convolution, reducing the parameter quantity of the model to one-third of YOLOX-l. Then, we introduce the pyramid attention segmentation module and FReLU activation function in GhostNet to improve the backbone network’s ability to capture contextual information at different scales; The Py-PAFPN structure is also proposed by adding multi-scale pyramidal convolution to the FPN structure to efficiently fuse the image information in the output feature layer of the backbone network. Using Pascal VOC as the dataset, the experimental results show that the Ghost-YOLOX lightweight object detection model proposed in this paper has less number of parameters and computation, and achieves 89.44% accuracy, which is the best result in the comparison experiments.