Research On Image Recognition Algorithm Based On Deep Convolutional Neural Networks

Convolutional neural networks can learn recessive features,it has strong self-learning capabilities,and achieved excellent results in the field of image recognition.The development of deep convolutional neural networks has brought profound changes to the field of image recognition,it can extract deeper features that people cannot understand and further improve the accuracy of image recognition.However,at the same time as the number of network layers deepens,the training of the network also encounters many problems.The first is that there is still no perfect mathematical theory to guide the design of the model structure,and the second is that the optimization algorithm used in the network training process has defects,both are directly related to the training speed of the network model and the accuracy of the model on the test set.This paper will focus on the deep convolutional neural network model and back propagation optimization algorithm to establish a simple and efficient deep convolutional neural network model.Firstly,based on the VGGNet network model,the full convolution is used to replace the three full connected layers of the original network.At the same time,a convolutional layer is added and the network parameters are adjusted appropriately.The improved full convolutional VGGNet plus(FC-VGGNet-plus)model is designed.The time per iteration is shortened by 0.13 seconds on the CIFAR-10 data set.At the same time,the accuracy rate on the test set is also improved,with the same device conditions and iteration times,the accuracy rate increased from 82.33% to 83.45%.The stochastic gradient descent algorithm is a very wide optimization algorithm in back propagation.Its idea is simple and the operation speed is fast,but it has inherent flaws in the application of deep convolution neural network parameter training.For the local optimal problem and the gradient diffusion phenomenon in the stochastic gradient descent algorithm for back propagation,this paper studies the particle swarm optimization algorithm in the swarm optimization algorithm and applies the modified algorithm to the convolutional neural network parameter training.The PSO-CNN based on particle swarm optimization algorithm is proposed.Firstly,the particle swarm optimization algorithm is used to train the network parameters,and then a stochastic gradient descent algorithm is used to further optimize the network parameters.Thissolves the local optimization problem of the back-propagation,and the fast iteration of particle swarm optimization algorithm greatly shortens network training time,basically,after 10 iterations,the accuracy rate is about 30%,however,it takes hundreds of iterations without using the particle swarm algorithm.At the same time,the convergence process of the network initiated by the particle swarm optimization algorithm is smoother,and the fluctuation of the parameter value is reduced.Afterwards,this paper applies PSO-CNN to the FC-VGGNet-plus network model training designed in this paper,and the improved full convolutional network designed is further optimized.This model has faster training speed and good convergence.The effectiveness of PSO in large-scale deep convolutional neural networks is verified through experiments,and the stability of the FC-VGGNet-plus network designed in this paper when using different optimization algorithms is also verified.Finally,an image recognition system is designed based on the new network model completed in this paper.The image recognition test is performed on the system and good results are obtained.