Application Research Of Object Detection Based On Simplified Convolutioanl Neural Network

Classification and detection is two of the most basic tasks in computer vision,and it is also the premise of other complex computer visual tasks.Since the birth of the concept of computer vision,it has been the direction of people's efforts to obtain high performance in the field of object detection.Artificial neural network is an earlier proposed method for data processing.In the early days,the effect of applying artificial neural network to object detection task was not satisfactory because of the processor performance and the fully connected structure of neural network.In recent years,with the improvement of the performance of graphics processor and the accumulation of a large number of data,it is a feasible choice to use artificial neural network to carry out the object detection task in computer vision.After the emergence of convolution neural network,deep learning continues to achieve better performance in object detection tasks than ever recorded,therefore,the application of deep neural network to target detection has become a research hotspot.Based on the existing deep learning methods,this paper simplifies the network structure by means of a certain method,and takes into account the speed and accuracy to the maximum extent.In this paper,a first-order deep detection network with YOLO as backbone is proposed.It uses the operation of separable convolution.The proposed network has a good balance between speed and precision,and can be transplanted to run on embedded devices such as Android.The main research contents are as follows:1.This paper introduces the development course of target detection and the current research status,including the task of object detection,evaluation standard,main Open data set,and two kinds of algorithms used in target detection.2.Investigating a large number of literature,from the artificial neural network's proposing,the working mechanism to its development process to give a detailed introduction.This paper introduces and analyzes two kinds of deep convolution neural networks using object detection: one-tage network and two-stage network.And the advantages and disadvantages of each of them are analyzed.3.The back propagation algorithm,the main parametric tuning method of deep neural network,is deduced and analyzed by formula.At the same time,several gradient descent methods for parameter updating are introduced,and the gradient explosion and gradientdisappearance which appear in the process of parameter tuning using back propagation algorithm are deduced by formula.Several measures to solve or alleviate these phenomenons are listed.4.Introducing the main neural network compression methods including the separation convolution operation to be taken in this paper from their main idea and formulas derivation by consulting a large number of literature.5.Based on the 4 aspects mentioned earlier,a deep detection network with YOLO network as its backbone is proposed.The network uses a separable convolution module and predicting relative coordinates instead of abolute coordinates,while abandoning the last layer using a full-connection layer and replacing the full connection layer with multiple continuous standard convolution layers.In terms of datasets,the COCO DataSet and Pascal VOC DataSet are mixed,and the clustering method is used to determine the number and coordinates of the priori box.The parameter tuning is carried out using the different kind of gradient descent algorithm in a phased way.Experimental results show that the number of parameters of YOLO network is 56.24 M,and the number of parameters of network proposed in this paper is 7.8 M.The mAP tested by YOLO in the Pascal VOC DataSet is64.3%,and the network proposed in this paper is 51.1%.On the upper computer,inference speed of the compressed neural network is increased by twice times.Under the premise of greatly streamlining the number of parameters,while loss a little accuracy.Finally,with the Android platform app,the acquired network is deployed on the Android platform,enabling images to be obtained from the camera and targeted for target detection..