Long Multi-digit Number Recognition from Images Empowered by Deep Convolutional Neural Networks. (13th September 2021)

• Record Type: Journal Article
• Authors: Asif, Muhammad; Bin Ahmad, Maaz; Mushtaq, Shiza; Masood, Khalid; Mahmood, Toqeer; Ali Nagra, Arfan
• Is Part Of: Computer journal. Volume 65:Number 10(2022)
• Page Start: 2815
• Page End: 2827
• Publication Date: 2021-09-13
• Subjects: multi-digit recognition -- deep learning -- CNN -- MNIST -- SVHN; Computers -- Periodicals; 005.1
• DOI: 10.1093/comjnl/bxab117
• Languages: English
• ISSNs: 0010-4620

Abstract

Abstract: Scanning images and converting the scanned information into digital format is an active research area. Scanning is an automated, fast and efficient process as compared to the traditional data entry, and the resultant converted data is more accurate. Recognizing digits from the scanned images is a challenging task. To address this issue, most of the existing techniques perform multiple individual steps that are localization, segmentation and recognition. Some researchers also focused on adopting a unified approach that combined these three steps for multi-digit recognition of up to five digits. To cope with the modern requirements, a unified multi-digit recognition technique capable of recognizing more than five digits is the need of the hour. Considering this necessity, a unified multi-digit recognition approach is presented in the current study that can recognize sequences up to 18 digits long. The proposed technique is based on a deep convolutional neural network algorithm that performs two basic functions. First, it localizes and extracts the region of interest in the image, and then it performs multi-digit recognition. The proposed algorithm recognizes sequences of up to 18 characters that makes it one of the preferred recognition techniques among the existing algorithms. The proposed technique is compared with state-of-the-art techniques and is proved to be superior and robust. The experiments are performed on two datasets, and overall accuracy up to 98% is achieved.