Research And Implementation Of Object Detection System Based On Deep Learning

Object detection is one of the basic tasks of computer vision.The purpose of object detection is to find the object of interest in the image,meanwhile the position and type of the object need to be detected.With the development of deep learning,object detection algorithms using convolutional neural network have developed rapidly,and frameworks have been widely used in face recognition and other fields.However,current algorithms are still insufficient in determining the accuracy of object positions and categories,and the effective using of object bounding boxes and object relationships generated by network can improve the accuracy of both sides.This thesis focuses on selecting bounding boxes and make full use of relationships between objects.The main research contents are as follows:(1)A bounding box screening method based on intersection over union and clustering algorithm is proposed.In the model,the intersection over union value of bounding boxes will generate,which will be used as the basis for selecting boxes.Intersection over union value describes the accuracy of bounding box,with this clue added in the model,filtering boxed can be improved.After prediction a more accurate boxes,in order to improve the accuracy of boxes,in the bounding box screening process,a clustering algorithm is used to treat multiple bounding boxes of a same target object as a group of processing objects.The experimental results show that under the condition of using Faster R-CNN as the model basis,this method has a 3.2%improvement in accuracy compared with the existing bounding box screening methods.(2)An object detection architecture based on relation is proposed.By adding relation modules which is based on attention mechanism and can process appearance features of different objects to the network,detection network can get more information for prediction.The experimental results show that,under the condition based on Faster R-CNN,the accuracy rate can be improved by 2.8%after adding the object relationship module.(3)Designed and implemented a Client/Server architecture based vehicle detection application,carried out overall system design,database design and system detailed design,and completed functional and performance tests.