Improved Faster R-CNN Based On CSP-DPN

Abstract

For the feature extraction in deep-learning based model of object detection, although DenseNet has the strong ability of exploiting features, its learned feature may result in high redundancy. In contrast, ResNeXt can reduce the redundancy of feature effectively, however, it has the difficult to use high-level information to re-discover low-level features. In this paper, inspired by Dual Path Network (DPN), we proposed a new feature extraction strategy that make full use of DenseNet and ResNeXt to improve the feature extraction ability of original Faster R-CNN. In the feature extraction part, ResNeXt and DenseNet alternately take first half channels and second half channels of feature map to perform feature fusion, which can learn more comprehensive information and reduce model’s computational complexity. In addition, we also apply the Switchable Atrous Convolution (SAC) in Faster RCNN to enlarge receptive field, which can further improve the accuracy of the object detection. We evaluated our proposed model on coco2014 dataset, the experimental results show that we achieved about 9% higher of mAP than original DPN92-based Faster R-CNN.