Siamese-based offline word level writer identification in a reduced subspace

In this paper, we explore the notion of similarity learning by utilizing the Siamese neural network employing the residual framework for the purpose of writer identification based on offline handwritten input word images. Apart from being text-independent, our method does not impose any limitations on the number of characters of the input word image being employed also it can be used in real-world applications where input image patches with a few letters exist. The novelty in our proposal is in the exploration of a sparse-based model for representing the output feature vector of the Siamese network in a reduced dimensional space. We also formulate a divergence-based approach for assigning a saliency score to each component in the sparse representation based on their discriminatory power. The system efficacy has been demonstrated on well-known word-level databases and the results obtained are promising when compared with previous works.