Research Of Small Target Detection Based On Deep Convolutional Neural Network

With the wide application of deep learning,the technology of target detection and recognition is more and more applied to a lot of artificial intelligence systems.The intelligent monitoring system is an application scenario of this technology.At present,most of the target recognition models based on deep learning can greatly improve the accuracy of detection,but the computation speed is still not real-time.Moreover,these models are unable to detect the smaller objects accurately.Therefore,the problems arising from the detection and recognition methods of the above targets,it is significant application value to study the solution of the use of deep learning.First,this thesis discusses the existing problems of the target detection and recognition algorithm based on deep learning.This thesis mainly uses Faster RCNN as the research model,and uses this model to do some related test experiments.According to the experimental results,this thesis analyzes some key factors affecting the target detection and recognition algorithm.And this thesis the direction of improving optimization is proposed in this thesis.Second,the work principle and function of the basic structural layer in the convolution neural network are studied in detail by this thesis.For some of the best convolution neural network models,such as VGG,GoogleNet,ResNet and other models,this thesis also makes a detailed comparison and research.Through the comparative analysis,this thesis summarizes the network structure form which can achieve bettertrade-off the computing power of network features and network parameters.Then,in order to improve the detection speed as far as possible while affecting the accuracy of detection,the network structure of depth feature calculation is optimized.The network parameters are retrained to ensure the sensitivity of the network to the target features.The experimental results show that the network structure adopted in this thesis is sensitive to the target features in the image,and improves the receptive field while reducing the parameters of the model.Then,the sliding window method,selective search algorithm and RPN layer detection algorithm,which can be used in the convolution neural network,are studied.After comparing and analyzing the advantages and disadvantages between them,this thesis selects the network structure as the basic layer of detection.Then,through the study of the shortcomings of FasterRCNN in the use of regional network structure,this thesis proposes an optimization method for the detection algorithm.In this thesis,we use the method of region extraction on multi-scale feature map to detect,which not only reduces the number of priori detection frames,but also reduces the missed detection rate for small target detection.This thesis sets up the preparation work of positive and negative data samples before training and the training parameters of the network model.And through comparison experiment,this thesis proves that this model is more accurate than Faster RCNN and YOLO on standard data set.For small target detection,the detection effect of this model is better than the SSD model that can meet the actual needs.In real time,the detection speed of this model reaches 19 frames per second.