Object detection based on Online hard examples mining with residual network

Abstract

In order to detect objects more correctly in pictures, an object detection algorithm based on the Online hard example mining with residual network is put forward, which takes Faster R-CNN as a benchmark. The working method of Faster R-CNN based on deep learning is portrayed, and the drawback and improvement ways of the algorithm are analyzed. More specifically, a residual network is used to replace the original ZF or VGG network to obtain more effective deep convolution feature maps. Besides, in order to strengthen the generalization capacity of the learning network model, the network parameters with hard examples are regenerated during training to make the network training more effectively. Finally, results of the experiments on Pascal VOC2007, Pascal VOC2007+VOC2012 and BIT datasets indicate that compared with Faster R-CNN, our method improves detection accuracy by 3.5%, 7.1%, 6.4% severally on the three datasets.