Improved YOLOv2 Object Detection Model

Abstract

Aiming at the problem of the large number of model parameters and poor performance on the small-size object of the YOLOv2 object detection model, an improved YOLOv2 object detection model is proposed. Firstly, it improves the YOLOv2 by introducing depth-wise separable convolution replace the standard convolution used in the YOLOv2. The number of parameters on the convolution layer is reduced by 78.83%. Secondly, the Feature Pyramid Network is introduced into the detection model to replace the YOLOv2’s image feature fusion method and perform object detection tasks on multi-scale image features. As a result, the ability of the improved YOLOv2 detection model to detect the small-size object is enhanced. Experimental results on PASCAL VOC 2007 datasets show that the improved YOLOv2 has a competitive accuracy to YOLOv2 and better performance on the small-size object.