Research On Text Extraction And Inpainting Algorithms In Colored Images

Text embedded in colored images carries important semantic information, therefore the technology of text extraction and translation is very important for image understanding and content-based information retrieval systems. However, most embedded text is surrounded by complex background and sometimes accompanied by high noises, which have posed great challenges to text extraction, and also fixed text hindered the overlay of translation results, thus restricted the application of machine translation. In this paper, we proposed some algorithms to achieve text translation on images.Text extraction may employ different text features. Some classic techniques are presented as an extraction attempt. However, owing to complex situations, none of these methods can extract text effectively.Thus, we improved an edge based text extraction method. It is carried out by edge detection, connected components labeling, morphological operations and local thresholding. At last we synthesize all connected regions to get the extraction of text. The method is robust to various background complexities and text appearances. Several experimental and comparative results are also presented.To remove background efficiently, three classic image inpaiting methods are analyzed.. Most of them are complex and quite time consuming. To overcome the disadvantage, a cubic convolution interpolation algorithm is used to downscale the original image and to improve inpainting efficiency. Firstly, the dimension of the original image was downscaled by a factor0.2-0.5, then the inpainted point with the highest priority from a target region was evaluated and a best-exemplar was searched from the source region in the downscaled image. Furthermore, those counterparts were obtained by the proposed rules from the original image and were filled into the whole damaged region in the origin image. The above steps were iterated until the target regions were completely filled. Experimental results demonstrate that the proposed image inpainting algorithm can obtain good inpainting results in efficiency and quality.Besides, as an application of those algorithms, we remove the detected texts from the original images, meanwhile, restore the background occluded by texts in images, and then put the translation results on the background. Experiment results show that the translation performs fairly well.