Improved Region Proposal Network Based Object Detection

The goal of visual object detection in computer vision is to detect and locate objects of interest in images or videos,and to determine which class they belong to.Object detection is one of the fundamental problems in computer vision,which has great value in both theory and application.However,current methods of object detection can not meet the requirements of tasks in practice in terms of accuracy or efficiency.That is to say,object detection remains to be an important and challenging research topic.Traditional object detection methods suffer from poor accuracy and low efficiency.The main reason lies in the insufficient quality of features extracted by manually designed filters.Deep convolutional neural network relieved this to some extent,which can extract feature maps from data automatically and efficiently in an end-to-end manner.Thus it has been broadly used in object detection.Object detection methods can be roughly divided into two categories.One is to treat object localization and classification as two subtasks,using different branches of a network respectively;the other uses an end-to-end network to accomplish both tasks simultaneously.This thesis proposed a few improvements on the first kind method,where its performance of object detection increases especially when there are multiple scale objects.The accuracy of the object detection algorithm using deep convolutional neural networks depends on the quality of image feature extraction to a large extent.Most of the existing feature extraction methods only consider the object of a certain scale,which is insufficient for multi-scale object detection,especially when there are small-scale objects.In this thesis,we proposed to use dilated convolution to construct a top-level feature map with multi-scale receptive fields.According to the scale range of the object in dataset,a plurality of dilation ratios are set.So the feature extraction network can obtain the multi-scale feature map.Such a region proposal network branch and a object classification network can utilize the feature map with multi-scale receptive fields to better deal with different scales of interest objects,so as to improve the accuracy of region proposals and the accuracy of object classification.In addition,since high-level maps have high-level semantic features and larger receptive fields,it is easy to miss small-scale objects if use them to generate object candidates.We proposed to add a new branch of region proposal network to the intermediate convolutional layer,which is used to generate more small-scale object candidate.The intermediate convolutional layer has a smaller receptive field while retaining more image details,so it is more appropriate to generate small object candidates according to feature image in this layer.By using two region proposal networks,the proposed model is able to generate object candidates corresponding to different scales on the feature maps of different receptive fields,thereby ensuring the recall rate of object candidates.Since the image features are shared in all networks,the proposed object detection method improves the accuracy of the detection while keeping the efficiency of the computation.This thesis shows the positive effects of feature maps with multi-scale receptive fields on object detection accuracy.In the meanwhile,region proposal network for object localization plays an important role in the whole network,whose performance also influences the results of object classification to some extent.The proposed method can improve the accuracy of object localization without significantly increasing the computation complexity.