Abstract: Currently,in order to achieve high performance,most image classification methods require adequate training and learning process.However,problems such as scarcity of training samples and overfitting of parameters are often encountered.To avoid these problems,we presented a non-parameter learning algorithm under the framework of Naive-Bayes Nearest-Neighbor (NBNN),where non-negative sparse coding,low rank and sparse decomposition and collaboration representation are jointly employed.Firstly,non-negative sparse coding combined with max pooling is introduced to represent images,and local feature matrices of similar training image sets with low-rank characteristic are generated.Secondly,two kinds of visual dictionary with category labels are constructed by leveraging low rank and sparse decomposition to make full use of the correlation and diversity of images with the same category label.Lastly,test images are represented based on collaboration representation for classification.Experimental results demonstrate effectiveness of the proposed algorithm.

Key words: Image classification,Bag of visual words,Sparse coding,Low rank and sparse decomposition,Collaborative representation