Research And Application Of Image Classification Algorithms Based On Convolution Neural Network

Image classification algorithms provide people with strong technical support for object recognition and information management,and have been widely used in the fields of transportation,safety,medicine and so on.Most traditional image classification algorithms rely on manually extracting low-level features,such as shape,color,etc.,but often some objects that are difficult to distinguish will consume a lot of human resources.Convolutional neural network(CNN)is one of the efficient algorithms of deep learning(DL)in the field of image classification,but image classification algorithms based on CNN have shortcomings,such as excessive internal calculations,unstable optimization algorithms responsible for parameter updates,and low classification accuracy for complex scenes.Therefore,image classification algorithms based on CNN to improve the above shortcomings are studied in this paper.The main work of this paper is as follows:1.An image classification algorithm based on improved Mobile Net is proposed.To deal with the disadvantages of CNN,such as incomplete extraction of feature information,too large model parameter scale and low classification accuracy on datasets containing complex image features,this paper proposes an improved image classification algorithm(L-Mobile Net algorithm)based on Mobile Net neural network optimization strategy.Firstly,The method adopts a new form of depth-separable convolution,and derives the feature information of the previous layer,and the feature information output by the two channels is transferred to the next layer through the deep convolutional fusion layer;secondly,Leaky Re LU is introduced into all network layers in order to preserve more comprehensive positive and negative feature information,and a residual-like structure is added to avoid gradient dispersion;finally,the L-Mobile Net algorithm is compared with the other six algorithms on four datasets,and the classification performance of the L-Mobile Net algorithm is more significant from the comparison experiment results.2.An improved activation function based on depth-separable CNN is proposed.To overcome the disadvantage of low learning efficiency in the non-linear feature directions when CNN algorithms are used for image classification,this paper proposes an improved activation function SSR based on depth-separable CNN.The activation function organically combines the advantages of Re LU and Soft Sign activation functions,and introduces an updatable value a_d.Experimental results show that through four different depth-separable network structures,compared with the other five activation functions,the depth-separable CNN combined with the SSR activation function is more prominent in image classification performance,and highest accuracy rates reach 93.6%and 95.7%respectively.3.An improved optimization algorithm based on Nadam is proposed.To resolve the problem of large difference in effect when the regularization methods and optimization algorithms are used in CNN to solve the over-fitting phenomenon,this paper firstly discusses the equivalence and non-equivalence of L2 norm regularization and weight decay regularization under SGD and Adam respectively,and then an improved optimization algorithm that integrates weight decay regularization into Nadam is proposed,namely WD-Na algorithm.To prevent over-fitting phenomenon,the first-order moment estimation and the second-order moment estimation of the gradient during the parameter update process are biased and normalized,and the weight decay regularization is added in this method.Finally,the comparison of Nadam algorithm with other four optimization algorithms on datasets,and the classification performance of the CNN image classification algorithm with WD-Na is more effective from the comparison experiment results,and accuracy rates reach 95.79%and 75.12%respectively.