Discrimination between Handwritten and Computer-Generated Texts using a Distribution of Patch-Wise Font Features

Recently developed deep learning techniques allow us to generate images of handwritten-like texts that closely resemble a target writer's actual handwriting. Although these models are applicable to useful systems (e.g. communication tools for hand-impaired people), they also could be misused for document forgery by a malicious user. To cope with this problem, in this paper, we propose a text-independent method for discriminating between the computer-generated texts (CGTs) and actual handwritten texts (HWTs). Our proposed method only takes a single text image and recognizes whether the image is a CGT or a HWT. Characters in HWT images have various shapes even when the same writer writes the same sentence. This property is difficult to perfectly mimic even by state-of-the-art CGT generation methods. To capture this difference between HWTs and CGTs, we use the distribution of patch-wise font features. The proposed procedure for discriminating HWTs and CGTs is as follows: First, we divide a given text image into several patches and classify each patch into one of pre-determined standard font classes. Then we compute the histogram of the standard fonts, which is finally fed into the recognizer that discriminates between HWTs and CGTs. In our experiments, the proposed method achieved more than 96% of discrimination accuracy, which demonstrates the effectiveness of the proposed method.