NEOCR: A Configurable Dataset for Natural Image Text Recognition

摘要

Recently growing attention has been paid to recognizing text in natural images. Natural image text OCR is far more complex than OCR in scanned documents. Text in real world environments appears in arbitrary colors, font sizes and font types, often affected by perspective distortion, lighting effects, textures or occlusion. Currently there are no datasets publicly available which cover all aspects of natural image OCR. We propose a comprehensive well-a.nnot.uted configurable dataset for optical character recognition in natural images for the evaluation and comparison of approaches tackling with natural image text OCR,. Based on the rich annotations of the proposed NKOCR dataset new and more precise evaluations are now possible, which give more detailed information on where improvements are most required in natural image text OCR.