The Study Of Setting Region Proposals Of Object Detection Network SSD

Object detection is one of the most advanced directions in computer vision,which involves objects detection and identification in images.Nowadays,Single Shot multibox Detector(SSD)is the state-of-art detection network,which evaluates a small set of region proposals of different aspect ratios at each location in several features maps with different scales.Different from the pipelined detection algorithm,the network SSD achieves end-to-end detection in order to improve the speed.The setting of region proposals is closely related to the accuracy of the network,because it will guide the direction of network to learn the feature.As a generic detector,SSD employs a large pool of region proposals discretizing the output space.However,for a specific dataset,the setting of region proposals may not fit.Hence,it is necessary to set suitable region proposals according to the dataset,which is vital for getting better performance.Aiming at these problems,this thesis mainly studies the setting of region proposals of SSD.Through theoretical analysis and experiments verification,this thesis explores how to set the region proposals according to the distribution of the dataset.Analyzing the influence of region proposals on performance of detection,this thesis gives the guideline of setting region proposals according to the distribution of dataset,size and shape of bounding boxes.Specifically speaking,too few species lead to too much difference of training samples for the regression mechanism learning of region proposals,leading to poor relationship and lower accuracy.While too much will bring interference to the regression relationship learning of the adjacent region proposals,which makes accuracy decrease.And with the increase of parameters,the complexity of model is increased,and the optimal solution is difficult to obtain.Compared with the original model,the modified SSD model improves both precision and recall,and reduces parameters.Experiments show that resetting suitable region proposals can avoid interference of background noise,and the detection accuracy of the network is greatly improved after appropriately increasing the intersection over union(IOU)matching threshold.