Convolutional Matching Technique for Urdu Text Recognition

Visual understanding of linguistic information is a crucial research topic because of its applications like Optical Character Recognition (OCR), opinion mining, and language translation. This paper focuses on the automatic Urdu text recognition system. Our work primarily consists of Urdu text recognition using a convolutional template matching technique. To this end, this paper investigates the character-wise explicit Urdu character detection and recognition method with Nastaleeq font. The proposed method comprises the grouping of homogeneous shaped classes that have their diacritic marks identified independently in a synthesis manner. Furthermore, we conduct experiments using an advanced Urdu News Ticker (UNT) dataset that provides precise character component-level labeling. The presented method attains 82.33% on the Urdu Printed Text Images (UPTI) dataset.