Research And Application Of Image Classification Model Based On Convolutional Neural Network

As a basic problem of machine learning and pattern recognition,image classification aims to study how to accomplish large scale image classification efficiently,which is of great theoretical significance and practical value in the present and future.In recent years,deep learning has been developing rapidly in the field of image classification.As one of the most important architectures in deep learning,convolution neural network is widely used in image classification.it simplifies network model by using weights sharing,local connection and sub sampling operations,which effectively reduces network training parameters and improves training efficiency.Due to the supervised learning algorithm used in convolutional neural network training,there will be problems such as insufficient training samples and gradient diffusion.So in response to these problems,this paper combines unsupervised learning algorithm with convolution in convolutional neural network,and KDL image classification model is proposed,which is applied to the classification of large-scale image data.Moreover,in order to solve the problem of excessive training parameters caused by the large scale of convolutional neural networks,a fuzzy DSD algorithm is proposed based on DSD algorithm and applied to convolutional neural network to improve network performance.The specific research content of this article includes the following aspects:1.K-means are combined with convolution into KDL algorithm.The K-means algorithm is used to extract the dictionary(clustering centers)of the input image blocks,then the corresponding feature extraction functions are constructed according to the dictionary,the final image features are extracted by convolution operation and sent to the classifier for classification.Experimental results show that the KDL algorithm is superior to some classical algorithms in image classification performance.2.Fuzzy DSD algorithm is proposed.Firstly,the Dense phase training is carried out,and the trained network weights is taken as the initial weights of Fuzzy training phase.and in the Fuzzy training phase,the network weight value is modified according to their proportion in each level of the network,then save the trained weights in the fuzzy phase as the initial weights of the next phase training;finally,the learning rate is adjusted and Dense training phase is carried out again.The fuzzy DSD algorithm is tested on the dataset.Experimental results show that the performance of Fuzzy DSD algorithm is better than DSD algorithm.3.Vehicle type recognition application.Firstly,The input vehicle images is preprocessed,including extracting image blocks of the same size,normalization,whitening processing,etc.Then K-means and convolution is combined to extract image features and use softmax to classify final image feature.The KDL algorithm is tested on the vehicle dataset,experimental results show that the proposed algorithm has less training parameters and higher accuracy than other algorithms.