Research On Text Location In Natural Scene Images Based On MLP And Region Analysis

With the rapid advances in digital technology, the extraction of text information from color images has gained increasing attention in recent years. Text information appears in images can be important in indexing and compressing of the images. Document image processing and understanding has been extensively and deeply studied over the past 40 years. Work in the field covers many different areas including preprocessing, physical and logical layout analysis, OCR/ICR, graphics analysis, signature verification, and writer identification and has been applied in numerous domains, including office automation, digital libraries, etc..In general, text in images can be divided into artificial text and scene text according to the existent form of text. Traditionally, document images are got from scanned paper, and they are mostly artificial text. There've been numerous works on artificial text. Recently, people has seen an increasing interest in adapting digital cameras and digital camcorders to get document images or scene images. So, a new question is presented for text extraction.In this thesis, we study on location of natural scene images. It includes five parts: pre-processing, feature extraction, classification, candidate text region generation, candidate text region analysis.In the stages of feature extraction and classification, we get the"ç±³"region in square window as the input features, and use MLP network to differentiate between text and non-text pixels. This approach can avoid the complex stage of constructing and choosing the features.In the stage of candidate text region generation, we proposed projection method to generate candidate text regions Based on the characteristics of the binary image got from MLP networks. In contrast with traditional approach using connected component method, this approach can avoid regions overlapping or overlay, and avoid the generating of small false text regions, and decrease the number of redundant text regions, and simplify the post-processing stage. In the stage of candidate text region analysis, we propose the approach of eliminating non-text regions using frequency information. Experimental results show that, this approach increases the location precision effectively.We determine property of each pixel based on the output of the MLP networks'outputs, then candidate text regions are generated using projection method based on results of the previous stage, finally, non-text regions are eliminated by candidate text regions analysis. Experimental results show that, our approach can get good text location result and reach an ideal evaluation criteria.