Keyword Spotting in Historical Bangla Handwritten Document Image Using CNN

In this paper, we propose a segmentation-free keyword spotting method for historical Bangla Handwritten Document images. For a given query keyword, the proposed method generates a set of candidate text regions within a document image using Histogram of Oriented Gradient (HOG) features. Next, the set of candidate text regions are fed to a trained Convolution Neural Network (CNN) model. The network finds the class label of each candidate text region and the regions whose class labels are same as the class label of the query word, are spotted in the document image. The proposed method is tested on three historical handwritten Bengali datasets and one historical English handwritten dataset. Detailed performance analysis demonstrates the goodness of the proposed method.