Applying Non-negative Matrix Factorization to Classify Superimposed Handwritten Digits

Abstract

Computation models that rely on a decision boundary approach have a severe limitation in recognizing the same objects with different appearances. Variations in the appearance of an object often create variations in their discriminative features as well. Hence, the derived feature data points might be shifted into the wrong decision regions. In this paper, we present an application of a dictionary-based Non-negative Matrix Factorization (NMF) for classifying handwritten digits under superimposition. The super- imposed digits are encoded as the input V, which is factorized into two matrices, V ≈ W H where W is the basis vector representing the prototypes of handwritten digits and H is the encoding coefficients representing the digit classification output sequence. We show that a dictionary-based NMF could classify an input sequence of single handwritten digits as well as superimposed digits. This is because NMF searches for an optimal linear aggregation of the components in the dictionary V ≈ W H. This unique abil- ity of NMF allows it to classify superimposed digits. The experimental results show that the proposed approach can recognize superimposed digits with good accuracy.