Deep Learning-Driven Multi-Stage Text Recognition in Financial Documents: Overcoming Overlaps and Faded Handwriting in Bank Cheques

Handwritten bank cheque images (BCI) can be segmented to extract crucial information and improve transactions. However, the segmentation process was disrupted in earlier studies when overlapped Handwritten Texts (HT) and Printed Texts (PT) were not concentrated. As a result, this framework successfully eliminated the overlapped texts for exact HT segmentation using Intelligent Character Recognition (ICR) and Pseudo Letter with Height-based Segmentation (PLHS). Initially, a denoised improved image is obtained by preprocessing the BCI. The overlapped HT with PTs from the improved BCI are identified and segmented independently using PHLS and ICR. After that, Nanonets are used to independently identify the date, signature, name, and amount from HT. The faded texts are restored using contour construction, edge detection, and texture inpainting. Finally, the recognized texts are categorized as forged and genuine cheques using a Sigmoidal Growing Cosine Intermap Pooling-based Convolutional Neural Network (SGCIP-CNN) with a classification accuracy of 98.7899% for effective money transactions. Thus, the proposed work outperformed the existing methodologies.