Rectification of Curved Scene Text Based on B-Spline Curve Fitting
Kiptanui Linus, C. J. Prabhakar, S R Shrinivasa · Indian Journal of Science and Technology · 2024
Objectives: In this study, we proposed suitable technique for rectification of curved scene text which is followed by recognition of rectified text in order to improve the accuracy of the existing techniques. Methods: In order to rectify curved text, initially, we perform curved text detection using Look More Than Twice (LOMT) model which detects and locates curved text. The detected text area is binarized through adaptive binarizaton technique. Then, we rectify the detected curved text through B-spline based curve fitting which align the curved text into straight line. The rectified text is feed to our recognition module where, we segment the rectified text using Low Variation Extremal Regions (ER) technique and extract the Pyramid Histogram of Oriented Gradients (PHOG) features from the segmented texts. Finally, we perform recognition of text using Tesseract OCR. The rectification and recognition performance of the proposed method on CUTE80 dataset was evaluated using Mean Square Error (MSE) metric and word level recognition accuracy respectively. We compared the recognition results of the proposed method with state-of-the-art methods using challenging benchmarks such as ICDAR2013, Street View Text (SVT), IIIT5k-Words (IIITK), ICDAR2015, SVT-Perspective (SVTP) and CUTE80 dataset based on word level recognition accuracy. Findings: From the experimental results, it is observed that the proposed rectification method removes highest error (MSE is minimum) of 99.34% for perspective text with large angle. Also, the proposed recognition module achieves highest recognition accuracy of 93.80% compared to the state-of-the-art methods on the selected six benchmarks. Novelty: We proposed a technique for rectifying curved scene text using B-spline curve-fitting technique followed by recognition using handcrafted features. The novelty of the proposed method is that we employ B-spline curve fitting in order to compute local transformation of individual character followed by rectification of each individual character through the computation of normal vectors. The recognition of rectified text is done through the handcrafted features. Keywords: Rectification, Curved Scene Text, Text Detection, Curve Fitting, Recognition, BSpline