The Research On Detection And Recognition Algorithm Of Small And Medium-sized Targets Based On Convolutional Neural Network

Target detection is an important research direction in the field of computer vision,which has important application value in self-driving,video surveillance and other fields.However,for a long time,limited by various technologies,the development of target detection is slow.In recent years,with its good generalization ability and excellent feature representation ability,convolution neural network has pushed target detection into the field of vision of the public again.In this paper,the target detection algorithm based on convolution neural network is studied.firstly,the basic theory of convolution neural network is briefly introduced,and then the principle,advantages and disadvantages of mainstream target detection algorithms are analyzed.Then,in order to solve the problem that the detection accuracy of Tiny-YOLO algorithm is not high,and it is easy to misdetect and miss small and medium-sized targets,an improved algorithm named Mini-YOLO is proposed.The structure of Mini-YOLO model is divided into two parts: feature extraction network and target detection network.In the part of feature extraction,this paper modifies the backbone network of the Tiny-YOLO model,so that it can extract the features of the input image better,and improve the utilization of features,then improve the detection accuracy of the model.In addition,a feature fusion layer is added to the feature extraction network to improve the recognition rate of the model for small and medium-sized targets.In the part of target detection,this paper briefly introduces the detection process,and explains the working principle of anchor box and NMS algorithm in detail.Then through training and testing the two models on the VOC2007 data set,analyzing and comparing the performance evaluation index,it is verified that Mini-YOLO can effectively improve the missed detection rate,error detection rate and the recognition rate of small and medium-sized targets.In order to solve the problem of vehicle and pedestrian detection in real scene,a vehicle and pedestrian detection network based on Mini-YOLO is designed in this paper.First of all,the method of data augmentation is used to solve the problem of imbalance of samples in the training set,and the new samples are labeled,so as to build an effective pedestrian and vehicle data set.Then the ground truths are clustered by k-means algorithm to determine the size of the anchor box.Finally,by selecting the pictures in the real scene for testing,it is verified that the vehicle and pedestrian detection network based on Mini-YOLO in this paper is effective and feasible,and can be put into practical engineering applications.