Research On Image Classification Based On Sparse Deep Learning

Image classification technology is the foundation of image processing and the improvement of image classification accuracy is very important to the development of machine vision.Aiming at the problem of low accuracy of image classification,a new method of image classification is developed by the improvement of sparse coding and the establishment of sparse deep learning model.The main results of this paper are as follows:1)Unsupervised deep spatial feature learning for aerial image classification.Aiming at the problem that the correlation and the spatial characteristics of the image feature extracted by the traditional sparse coding method is not obvious,a new algorithm is proposed.The local variance similarity is introduced into the traditional sparse coding and sparse representation of the extracted image features is used to enhance the correlation and the spatial characteristics of the image features.Finally,combined with deep belief network,the aerial image classification experiment is completed.2)Unsupervised deep discriminative feature learning for aerial image classification.In order to improve the discrimination of the non negative sparse coding,a new method based on unsupervised deep discriminative feature learning is proposed.The Fisher discriminative analysis criterion is introduced into the non negative sparse coding,which improves the discrimination of the image feature and lays the foundation for improving the classification accuracy.3)Natural image classification method based on sparse deep belief network.In order to enhance the sparsity of RBM and build a sparse deep belief network model,this paper proposes to use hyperbolic tangent function to approximate 0L norm,which can be made as the sparse RBM penalty term.Proposed algorithm obtains high classification accuracy on Natural image classification experiments.Improved algorithm can effectively extract the image features and has a good classification effect in aerial images and natural images,which provides a new way to improve the accuracy of image classification.