Research Of Object Detetion Based On Multi-scale Convolutional Features

Object detection is one of the basic tasks in computer vision,and the efficient target detection algorithm has always been pursued by researchers.From the perspective of multi-scale features in deep learning,we studies multi-feature fusion methods and combines the high-order features with location aware weighted networks.Firstly,we study the multi-feature fusion method based on Faster-RCNN and feature pyramid network(FPN),respectively.It is necessary to study these two methods because the Faster-RCNN based method has only one prediction component while the FPN based method has multiple components.Two multi-layer feature fusion strategies are proposed in this paper and verified on the PASCAL VOC dataset,and effectively improve the accuracy of object detection.In addition to the study on multi-layer feature fusion strategy,we further propose a multi-scale location-aware kernel representation method(MLKP).MLKP proposes an application method of high-order features in the filed of object detection and designs a location aware weighted network and combines them together.The use of high-order features is a feature-enhancing method that can improve the ability of models to detect small objects.By using the statistical information of the target location,the location weighted network can perform weighted operations on the feature map.In the feature map,not all areas will have targets.The location weighted network can enhance the area where the target exists,weaken the area where the target does not exist and achieve the purpose of improving the target detection effect.In the meantime,we study the multi-scale location weighted network and analyze its advantages and disadvantages.Finally,the MLKP is validated on two types of published datasets: PASCAL VOC and MS COCO,and has achieved the best current results.