Utilization of information from CNN feature maps for offline word-level writer identification

Abstract

This paper proposes a text-independent method for authorship identification using handwritten word images. Our method is text-independent and imposes no limitations on the size of the input word images being analyzed. To begin with, the SIFT algorithm is utilized to extract key-point regions (fragments) across different levels of abstraction, encompassing allographs, characters, or character combinations. These fragments are subsequently processed through a CNN network, yielding feature maps corresponding to convolution layers. The information in these maps is then transformed into a fixed-dimension representation using a modified version of the HOG feature descriptor.The noteworthy contribution of our proposal lies in harnessing additional cues from CNN’s feature maps for writer identification. We propose a measure to gauge the importance or ‘saliency’ of feature maps within a CNN layer during training. This measure originates from applying Sparse Principal Component Analysis (SPCA) to Histogram Of Gradient (HOG) features. Once saliency values are obtained, we combine them with HOG representations to create writer descriptors tailored to a CNN layer. The derived descriptors are then passed through a set of SVM classifiers that are scored at two levels to determine the identity of the handwritten word image.The efficacy of our system has been demonstrated on the word-level data of the CVL, IAM and CERUG-EN datasets with an accuracy of 82.10%, 87.68%, and 75.80% respectively. The results obtained are shown to be promising when compared with previous works.