GFSPP-YOLO: A Light YOLO Model Based on Group Fast Spatial Pyramid Pooling
Shaojie Xu, Yujiao Ji, Guangcheng Wang, Lei Jin, Han Wang
Abstract
The YOLO object detection model for PC environments is widely used in computer vision due to its high accuracy and good real-time performance. However, when faced with the embedded environment of mobile devices, the use of YOLO models in mobile devices is still challenging due to the large computational requirements and memory consumption. To address these issues, this paper proposes a lightweight YOLO model based on grouped fast spatial pyramidal pooling. Different from the existing YOLOv5 model, firstly, at the end of the backbone network, the receptive field is expanded using the ideas of CSPNet and group convolution to build a group fast spatial pyramidal pooling structure GFSPP to avoid false and missed detections caused by image distortion; and a CBAM attention mechanism is introduced in the backbone network to improve the characterization of network features. Secondly, the slim neck paradigm combined with the lightweight convolutional module GhostConv is used in the neck network to compress the network structure. Finally, migration learning techniques are used to further improve the detection performance of the model. Experimental results show that the GFSPP-YOLO model proposed in this paper reduces the complexity and parameter costs by 10% and 3.5% respectively compared to the traditional YOLOv5s model on the PASCAL VOC2007+12 dataset, while the mAP0.5 is improved by 2%, making the model in this paper more suitable for applications in embedded environments of mobile terminals.