Research On Target Detection Algorithm Based On Deep Learning

Target detection in the field of computer vision is an important part of the basis of computer vision is an integrated subject in the field of computer science and mathematics and so on.Through the input of images and videos,the machine analyzes and processes it,and then extracts the information contained in it,and finally realizes the function similar to human vision.With the continuous development of electronic information and Internet technology,in some specific environments and devices,important information in the image needs to be extracted,and some interested targets need to be detected and identified.There are two main tasks in target detection,namely object classification and location.It is to screen the objects of interest in complex scenes and determine their categories and specific positions.Because the size,attitude,color and other characteristics of different things are different,it is very challenging for the target detection technology to be applicable to different scenes.This article through to the target detection algorithm based on depth of learning are studied,in order to improve the detection accuracy,the target and small targets detection difficult problems,such as,the Faster R-CNN network model to improve the optimization,through the experimental simulation,this paper to improve the target detection algorithm is compared with the original Faster R-CNN algorithm for target detection performance is more superior.The specific research contents of this paper are as follows :1.Improve the accuracy of target detection.First of all,the structure of the Faster R-CNN algorithm is analyzed.The Faster R-CNN algorithm is composed of a convolutional neural network,a regional proposal network(RPN),and a Fast R-CNN network.The algorithm uses a convolutional neural network for regional selection,rather than using another algorithm for separate operation.Among them,RPN and Fast R-CNN share the convolutional neural network for feature extraction,which reduces the number of convolution computation and improves the speed of the whole algorithm.In order to improve the accuracy of target detection,a visual attention mechanism is introduced.Visual attention mechanism enables people to quickly locate the region of interest and find the target they are interested in when facing a complex scene,and selectively ignore the region of no interest.Through the combination of visual attention mechanism and RPN network,the saliency of the target in the image can be improved,so that the accuracy of target detection can be improved in the final target detection.2.It is difficult to detect small targets.In the process of target detection,some small-scale targets are difficult to detect because of the different target scales in the image.Therefore,in order to solve this problem,this article through to Faster R-CNN algorithm of feature extraction network is analyzed,by using characteristics of the Feature Pyramid Networks(FPN).A multi-feature fusion strategy for feature extraction is proposed.Feature extraction figure will step by step to reduce the network every time the output characteristics,the features of the characteristics of each layer of the output information is different,the characteristics of the deep information will ignore some of the underlying edge,color,shape,texture and so on low level information.Based on feature extraction on each layer to the output the network down and bottom-up integration,then two fusion by pressing the element after the formation of the characteristics of the pyramid sum fusion is way,so that the final output features contain not only the semantic information of deep feature layer,but also the feature detail information of shallow feature layer,which improves the detection accuracy of the algorithm for small targets accuracy.