Structure Design Of Convolutional Neural Network And Its Application In Target Detection

Since it was proposed in 2006,deep learning has attracted great attention from academia and industry,and has quickly become the most active branch in the field of machine learning.Convolutional Neural Network is a very important network in the field of deep learning,which is widely used in many fields and has made outstanding achievements,especially in computer vision,where it has made continuous breakthroughs in image classification and target detection.Convolutional neural network architecture is simple,but it can achieve more efficient network training through the final combination of different modules.Convolutional neural network training does not need to actively extract image features,and can directly carry out end-to-end training and prediction,breaking the traditional algorithm image preprocessing process.Recently,convolutional neural network structure innovation and parameter optimization have made some achievements in computer vision related tasks,but with the increase of network depth and network model,it is easy to appear over fitting,and it is difficult to be applied to embedded hardware devices.Therefore,starting from the basic architecture of convolutional neural network,this paper aims to design a convolutional neural network with superior performance in all aspects,and reasonably use the algorithm of convolutional neural network to apply it to embedded devices.The classical convolution neural network has single convolution operation,single convolution module and long training time.It needs to design a variety of convolution modules for complex convolution to break the tradition and improve the performance;In this paper,the application part uses GSM module,through the TCP serial communication between the computer and MCU,and combines convolution neural network algorithm with embedded hardware such as MCU to apply to target detection,and finally realizes the design of home security monitoring system based on convolution neural network algorithm.The main research contents of this paper include the following aspects:(1)the traditional convolutional neural network has a single network structure,insufficient feature extraction and large network model,so a parallel cross module Cross Input-Net is built instead of the traditional single convolution layer to extract image features,and two channels of convolution kernels with different numbers and sizes are used to expand the receptive field,while residual learning is added to deepen the network and avoid over-fitting.Two kinds of convolutional neural networks Net 1 and Net 2 are built,and the recognition performance is tested on two public data sets 101?food and GTSRB.the experimental results show that the recognition rate of the improved network is higher than that of the traditional neural network and the model parameters are greatly reduced.(2)The traditional network convolution method is single,the module is single,the overall network model is large,the training speed is slow,and the training cost is high.Therefore,a diversified convolution neural network is built,which uses Reduce moduel and Slice Module in combination,and adds Shuffle Net unit in the module to reduce the computational complexity of the model and improve the training speed of the network.Residual learning continues to be used in the network to reduce the over-fitting problem,and global mean pooling is used instead of full connection to reduce the model parameters.The constructed convolutional neural networks Wei Net-shuffle and Wei Net,as well as the previously published networks,have been tested on the public data set 101?food.The experimental results show that the parameters of the improved network model are much smaller than other networks,and the recognition rate is slightly higher than other networks,but the training speed of the network is obviously higher than that of the traditional network.(3)Household monitoring equipment has been widely used,but most of them are simply connected with the computer for monitoring,and the monitoring is called up only after the accident,which is not real-time,and the performance of simple hardware equipment is limited.It is difficult to quickly detect the target object in monitoring by combining convolutional neural network algorithm with embedded hardware platform.In this paper,based on YOLO V3,the camera first transmits the collected image to the computer through the router,and then the computer uses convolution neural network algorithm to detect the target of the collected image.When the computer detects the target,it runs the script program written by python,and transmits the instructions of the upper computer to the serial port of the MCU through the router through TCP communication.The MCU receives the corresponding instructions,and the program sends the SMS to the user's mobile phone through the GSM module.The user can open the SMS and click the video address to view the home monitoring in real time.