Scene Text Extraction from Natural Images Using a 2D Haar Wavelet Transform Pipeline

Abstract

Text extraction from natural scene images is a fundamental task in computer vision with applications in areas such as autonomous navigation, assistive reading for the visually impaired and real-time translation. Natural scenes pose serious challenges such as cluttered background, uneven illumination, perspective distortion and multiscale and multi-oriented text. In this paper, a new training-free signal-processing framework using the two-dimensional Haar wavelet transform is proposed for the scene text extraction. The pipeline consists of a luminance based gray scale conversion and an adaptive wiener filtering for noise suppression followed by haar wavelet decomposition into directional sub-bands. Sobel edge detection with sub-band fusion highlights the text boundaries. The morphological dilation operation is applied, then connected component analysis with a geometric density filter and Otsu binarization is performed to isolate the candidate regions for Tesseract recognition. On three standard benchmark datasets the method achieved F-measures of 89.5, 86.8 and 80.6 percent, which makes it competitive to both classical and deep-learning approaches while requiring no training data.