A CNN-Based Classification Model for Recognizing Visual Bengali Font

Automatic font recognition or similar font suggestions from an image or picture are the core design works for many designers. This paper proposes a framework based on Convolutional Neural Network (CNNs) to the widely neglected problem of Bangla font recognition by the vision community. First of all, we build up the available large-scale dataset consisting of both labeled synthetic data by Adobe and partly labeled real-world data. Next, CNN is trained to classify images into predefined font classes. Global average pooling layer is proposed instead of fully connected layers over feature maps in the classification layer to correspondence between feature maps and output. Thus, the feature maps can be easily interpreted as font categories confidence maps. We show that our method achieves state-of-the-art performance on a challenging dataset of 10 selected Bangla computer fonts with 96% line-level accuracy. Large-scale experiments show that our approach is exceptionally viable on our synthetic test images and achieves promising results on real-world test images.