Research On Object Detection Algorithm Based On Deep Learning

The task of object detection is to find and locate the objects of interest quickly and effectively from the input images or videos by computer,and give the categories of the objects.As one of the fundamental problems of computer vision,object detection is a hot and difficult problem,which has significant research and application value.The detection accuracy and speed which general detection task required could not be satisfied by current methods.And there are some problems,such as low detection accuracy for small-scale objects,poor positioning accuracy for objects,and insufficient ability to capture all kinds of objects at the same time.Therefore,this paper aims to solve these problems and improve the detection accuracy of the model by analyzing these problems.Firstly,this paper proposes an object detection model DF RCNN which uses fully convolutional networks to realize positioning branch to improve the detection accuracy of small-scale objects.It uses fast RCNN as the basic framework,adopts backbone network with better feature extraction effects,adopts multi-scale feature fusion strategy,constructs feature pyramid,enriches the semantic information contained in features,and makes them adapt to different objects with different scales.After that,ROI align technology is used to avoid the precision loss caused by pooling.At the same time,the fully convolution spatial information retention ability is used to replace the fully connection to form the object location branch.More accurate positioning of the object.At the same time,the sampling of the general convolution in object detection is fixed,and it is insufficient to detect flexible objects.This paper proposes a multi-scale object detector that combines deformable convolution and mask information.This detector improves the fixed convolution in the backbone network,combines the pyramid technology to achieve multi-scale information fusion,and further optimizes the candidate regions by using improved Soft NMS.Finally,the final result is obtained through the detection module,and the effective detection of the filter bag is realized.Finally,in order to verify the effectiveness of the above object detection models,this paper tests on Pascal voc207,Pascal voc2012 and the dust bags data set of vacuum cleaner constructed in this paper.The results show that DF RCNN can effectively detect small-scale objects,and the location results of objects are better than other models;Combining the deformable convolution with Mask RCNN can effectively detect the filter bag.