Research On Few-shot Object Detection Technology

Object detection is an important research direction in computer vision.Its goal is to ide ntify the position and category of objects in a given scene.At present,the deep learning me thod based on mass data annotation is the mainstream method of target detection,which is essentially based on mass data to extract features for target detection tasks.However,in so me applications,such as the discovery of rare targets in the military field,rapid screening o f suspect molecules,and monitoring of endangered organisms,it is relatively difficult to ob tain a large number of labeled samples.In these scenarios,how to get a model with good g eneralization ability in the case of few samples is a challenging problem in object detection field.In this paper,small sample target detection research is carried out from three directio ns of small sample data enhancement,model structure and platform building.The main wo rk is as follows:(1)In terms of data enhancement,a solution based on Sin GAN(Single Natural Image Generative Adversarial Networks)generative adversarial networks is proposed,which com bines the existing data enhancement methods.Firstly,with the help of generative adversari al network to generate more similar samples from a single sample,the final data set is gene rated by combining simulated samples with traditional data enhancement methods such as Mosaic and Mixup.Experiments show that the data enhancement strategy proposed in this paper can improve the efficiency of model detection.(2)In terms of model structure,based on the residual block structure of Dark Net-53,th e feature extraction network of YOLOv3(You Only Look Once V3),A scheme that integra tes Sk Net(Selective Kernel Networks)and CBAM(Convolutional Block Attention Modul e),which are characterized by adaptive receptive field and selective convolutional kernel,i s established.Firstly,Sk Net,an attention mechanism network based on convolutional kern el,is introduced into the residual block structure of the Dark Net-53 backbone network,and on this basis,samples of different sizes and scales are output to the feature pyramid structu re FPN in YOLOv3,so that neurons can adjust the size of their receptor field adaptively ac cording to the input of multi-scale feature map.This will further enhance the feature extrac tion and target detection of small samples.Secondly,for the residual block structure of Dar k Net-53,by adding CBAM module to the residual unit,the spatial and temporal dual-dime nsional parameters of sample features can be effectively adjusted,so as to improve the acc uracy and reliability of feature information extraction,and make the residual block networ k reduce the disappearance of gradient information and reduce the interference of irrelevan t information such as target background.Capture the more critical target information of the sample.Experiments show that this method can improve the generalization ability and pre cision index of the model.(3)On the basis of data enhancement and model structure,the visualization interface is constructed,and the small sample detection function for image and video is integrated.Th e whole system can be packaged into a general unit program independently,and the small s ample model target detection system is built,and the small sample target detection method proposed in this paper is verified on the experimental dataset.