Research On Object Detection Algorithm Based On Faster R-CNN

Object detection is an important task in the field of computer vision,and its main task is to detect the specific object from video or image.Object detection uses image processing,pattern recognition,artificial intelligence and so on.It has wide application prospect in military,video surveillance,human-computer interaction and so on.The traditional object detection algorithm has achieved good results in some simple scenes,but it is difficult to meet the demand in complex scenes.The rise of deep learning brings new ideas for object detection in complex scenes.With the powerful feature extraction ability brought by its deep level networks,deep neural networks abstract the low-level features of the original input data into high-level features,which is more conducive to object detection.The status of the object detection algorithm is first analyzed in the thesis.Then the Faster R-CNN(Faster Region-Based Convolutional Neural Networks)algorithm is studied.Finally we realize two improved algorithms based on the Faster R-CNN algorithm and verify them on the PASCAL VOC 2007 dataset.The main research work we have done in the thesis is as follows:(1)We study the network structure that RPN(Region Proposal Network)network and RCNN(Region-Based Convolutional Neural Networks)network has their own feature extraction network in the Faster R-CNN algorithm.The experimental result shows that the accuracy of the improved network structure is 2 percentage higher than that of the Faster R-CNN algorithm.(2)In Faster R-CNN algorithm,the quality of the proposals generated by RPN network is low and there are a lot of invalid proposals.In order to solve this problem,we optimize the proposals generated by the RPN network through a optimizing network,which makes a large number of invalid regions to be filtered.The experimental result shows that the accuracy of the improved algorithm and Faster R-CNN algorithm are almost the same,but the average detection time of each image of the improved algorithm is 7ms less than that of the Faster R-CNN algorithm,which proves that the improved method is effective.