Super-Resolution of Textual Images using Autoencoders for Text Identification

Textual image identification assumes a particular and sufficient resolution for the Region of Interest (ROI) on the image. The object or letter to be detected on the image must be of specific resolution otherwise most of the system fails to detect properly. Convolutional Neural Networks trained on normal textual images fail to detect characters, which is upscaled by interpolation methods. A deep neural network based model is proposed in this paper for upscaling and detection of textual images that show a significant increase in performance on the standard dataset. The proposed model consists of a customized autoencoder network for 4× upscaling of an image so that an automated system like the convolutional neural network can detect the upscaled images accurately. Results show a significant increase in the Structural Similarity Index (SSIM) and the Peak Signal to Noise Ratio (PSNR) measures for using the proposed model. Most importantly, the proposed approach achieves 96.44% identification accuracy for the Super-Resolution (SR) images.