Learning sparse non-negative features for object recognition

Vision based object recognition has attracted much interest in recent years due to its spread area of applications. Purely computer vision techniques, biologically motivated approaches or combined methods have been developed to tackle this task. Object recognition task based on three variants of non-negative matrix factorization techniques is investigated in this paper. The analysis is undertaken with respect to the recognition performances of the three investigated algorithms namely, non-negative matrix factorization, local matrix factorization and discriminant matrix factorization. The correlation between the sparseness of basis images and the classification accuracy is also considered. The experimental results reveal the fact that, the degree of sparseness is an important issue and differently affects the recognition performance for each algorithm.