Segmentation free text recognition for overlapping characters using spectral features and bidirectional recurrent wavelet neural network
Neha Goel Tripathi, Pushpinder Singh Patheja · International Journal of Intelligent Engineering Informatics · 2022
This paper addresses the problem of text recognition with overlapping and touching characters which is very challenging task due to its inability to be segmented effectively. A novel framework based on word level recognition has been presented in this work which does not require the character level segmentation. Dynamic window selection technique for different aspect ratios of input images is presented to overcome limitations of existing techniques whose performance is subject to the normalised window size and uniform aspect ratio prior to the feature extraction. The combination of discrete wavelet transform (DWT) and histogram of oriented gradients (HOG) is used to represent the characteristics of overlapping and touching characters. The robustness and accuracy in the recognition phase is enhanced through bidirectional adaptive recurrent Wavelet Neural Network (BRWNN). The word accuracy in the proposed work is achieved to be 83.14% which is better than the conventional techniques over the dataset IAM-DB.