Object Detection Technology Research Based On Multi-feature Fusion

With the development of the Internet and smart device,a large amount of visual data will be generated on the network.And how to make computer understand this visual data has become a difficult hotspot in academic and industrial research.This hotspot has promoted the development of computer vision and artificial intelligence.As an important fundamental branch task of computer vision,object detection technology is a predecessor of many advanced vision processing and analysis tasks.In the field of generic object detection,this thesis focuses on improving the detection performance of common object detection algorithms based on deep learning,and carries out a series of research work.Then,a novel object detector based on context modeling and multi-scale feature fusion scheme,CMMFD(Context Modeling and Multi-scale Fusion Detector),is proposed to improve the detection performance of commonly used object detector SSD.The main contributions of this thesis are as follows:Firstly,through in-depth analysis of the public object detection algorithm SSD,it is found that the existing public object detection algorithms cannot fully utilize the spatial information of images.Aiming at the common problem,this thesis designs a context modeling scheme based on dilated convolution and a multi-scale feature fusion scheme that fuses information at different scales.Secondly,by integrating the designed context modeling scheme and multi-scale feature fusion scheme into the SSD,a new object detector CMMFD is designed,and its performance on PASCAL VOC and MS COCO datasets was tested experimentally.Through experiments,it was found that CMMFD can obtain 80.5% and 77.0% mAP on the PASCAL VOC 2007 and PASCAL VOC 2012 test sets,respectively.And it can also obtain30.7% mAP on the MS COCO dataset.This shows that CMMFD can effectively improve the detection accuracy of object detector.At the same time,the ablation experiment also proves that context modeling and multi-scale feature fusion can also improve the detection accuracy of object detector.Thirdly,in order to further optimize the detection performance of CMMFD,this thesis designs a multi-scale detail context feature to make up for the lack of depth context information,thereby improving the detection performance of CMMFD.Fourthly,in order to further improve the detection accuracy of CMMFD,this thesis combines ensemble learning to design an ensemble scheme based on NMS(Nonmaximum Suppression)and an ensemble scheme based on feature map,and compares the performance of different ensemble schemes through experiments.Experiments show that the ensemble scheme based on NMS can improve the detection accuracy,but it will reduce the detection speed.The ensemble scheme based on feature map has a better trade-off between detection accuracy and detection speed.Therefore,the ensemble scheme based on feature map is effective for improving the performance of the object detector.