Research On Image Recognition Based On Diversified Convolutional Neural Networks

With the advancement of the era of big data and the rapid development of computer hardware,the application of artificial intelligence is more and more closely related to our lives.In the field of artificial intelligence,deep learning has become a research hotspot in this field,and has made breakthroughs in research results.Among them,deep learning in applications involves smart driving,smart medical,image processing,and voice processing.In the deep learning algorithm,the convolutional neural network is the most widely used,especially in image processing,and has outstanding performance in performance.It has become one of the mainstream algorithms in image processing algorithms.Compared with the traditional image processing method,the convolutional neural network takes a two-dimensional image as the input of the network,through the layer-by-layer propagation and calculation of the network,extracts the salient features in the image,and finally realizes the image recognition process through the classifier..This process can not only effectively reduce the difficulty of traditional methods in image preprocessing,but also reduce the subjective judgment of artificial selection of effective features,and achieve end-to-end recognition image mode.This paper mainly studies the new convolutional neural network structure published in recent years,redesigns the new network structure,and proposes a new network integration method.Designed to avoid the singularity of the classic network,it can realize the advantages of other network structures,and finally achieve better recognition and save resources.The main research contents and innovations of this paper are as follows:(1)This paper introduces the main content of convolutional neural network,including the working principle of single neuron and multi-layer perceptron.It also introduces the back propagation of network and the overall structure of the network.The classic network is taken as an example to introduce convolution.The computational flow of the neural network;a brief description of the deep learning framework Caffe used in this paper;an introduction to data preprocessing such as scale normalization,de-averaging and data enhancement.(2)Aiming at the single structure problem in the classical network that has been published,a multi-dimensional structure convolutional neural network is designed to be applied to image recognition.The diversified structure network uses two branches for feature extraction,one branch is the traditional CNN,and the other branch is added with the residual operation on the basis of the traditional CNN,and the cascade operation is performed before the next feature map is reduced.Combine two different network branches.The design of the diversified structural network was tested in the Food-101 and Caltech256 data sets.The cascaded network was compared with the two branch networks.The experimental results show that the network structure after the cascade is more than the two branches.A network is highly accurate.(3)Combined with the Inception module structure of GoogLenet network,the diversified structure network is improved and a dual-network module structure is designed.Firstly,the overall structure of the Inception module is analyzed in detail,and the idea of the Inception structure is used to improve the network of diverse structures.The model structure can greatly reduce the utilization of resources while combining the advantages of the diverse structure network.The model divides the network structure into two branches,using convolution kernels as two 3×3 and a convolution kernel as a 5×5 convolution layer.Finally,the network model is tested on the public dataset Food-101 and GTSRB data sets.The experimental results show that the design has a good effect.(4)In order to further improve the diversity of the network,and the problem of more network parameters.This paper firstly quotes an ultra-lightweight network structure module,which integrates the dual-branch network module structure through the de-induction function of the ultra-lightweight network module,so that the overall network increases the module diversity.Finally,the Food-101 and GTSRB data sets are verified.The network not only absorbs the advantages of ultra-lightweight parameters,but also inherits the diversity of the network structure,so that the accuracy and parameters of the overall network are obtained.optimization.