Research On Image Processing And Recognition Algorithm Based On Convolution Neural Network

Convolution neural network(CNN)is a kind of deep learning,which has become a popular research in image recognition.Its weight sharing can effectively reduce the complexity of the network structure,reducing the number of training weights.There is no traditional classification algorithm in the feature extraction and data recovery process.The model can directly input original images.This network structure in the image processing and recognition has a good performance.The main research direction of this paper is the image processing and recognition based on CNN.A CNN not only can complete image processing,no longer need traditional methods such as image semantic analysis and feature extraction,but also can integrate features into categories for identification.In recent years,the study of image processing and recognition based on CNN is a popular research field in computer vision,image processing and deep learning.Due to the rapid development of CNN,a large number of excellent regularization methods and other technologies have been introduced into the process of image processing and recognition.This paper is resolved problems in image processing and recognition based on CNN with the method of dropout and the simulated annealing algorithm.The main research results are as follows:1.For the problem of over-fitting and model averaging methods,dropout is that the part of activation units are suppressed and are likely to be activated in the next training.This method reduces the joint adaptability between neuron nodes,improves generalization ability and avoids the over-fitting problem.However,there are shortcomings that cannot be ignored.Model averaging method only considers the impact of retaining probability.In this paper,an improved dropout method is proposed,and a new method of model averaging is proposed.2.In order to obtain a more generalized network,an improved CNN combined with simulated annealing algorithm based on cross entropy loss function is proposed.Heuristic algorithm used to improve the performance of CNN is still rarely investigated.In this paper,the simulated annealing algorithm based on cross entropy loss function is used to optimize the CNN of dropout regularization.The weights of last layer are tuned,then,back propagation algorithm is used to update the weights of the front layers.Experiments show that the image recognition rate of CNN based on the simulated annealing algorithm has a good effect.It can get a better recognition rate and can be better to overcome the over-fitting problem.3.For the problem of long training time,this paper also proposed a CNN with adaptive learning rate algorithm in order to clipping convergence.The learning rate in traditional neural networks is a global constant.The large learning rate is not conducive to the minimum value of loss function.The small learning rate will consume long training time.A method of selecting learning rate based on loss function is proposed.The experimental results show that the method is effective and practical,and can be quickly convergent gradually to achieve the best.