The Research On Uyghur Text Localization In Complex Background Images

Text in complex background images often contains important information that is the key to image content recognition and scene understanding.Therefore,the text localization technology in complex images is an important research topic in image retrieval,human-computer interaction,pattern recognition and so on.At the same time,it has broad application prospects in intelligent shopping,automatic driving and document analysis.The text localization in an image refers to found the coordinates of the text in the image so that the text could be recognized.At present,most of the research on the text localization in complex background images focuses on Chinese and English,which is not good at minorities language text(such as Uyghur)localization tasks.In addition,due to the noise,illumination,blur and occlusion,as well as the text font size,color and style,making the text localization task is full of challenges.In particular,complex background images often contain a large number of elements(such as branches,grass,wave etc.)that are confused with Uygur characters.This paper focuses on the methods of Uyghur text localization in the complex background image.Some achievements are summarized as follows:1.Component Extraction with Multi-color-channel Enhanced MSERIn recent years,the candidate component extraction method based on Maximized Stable Extremal Regions(MSER)has been widely used.But the MSER is sensitive to low-resolution and blur,and it is easy to omit the text component candidate.In order to compensate for this defect,this paper presents a candidate component extraction method based on multi-color channel enhanced MSER,which includes two stages,the multi-color-channel MSER extracting and the rapid duplication removing.In the former phase,the MSER in each color channel can effectively improve the robustness of the image and the low resolution of the MSER,and obtain a highly recall.In the later phase,the candidate components extracted in the multiple color channels which extracted a large number of repetitions,an efficient duplication removing algorithm is used at this stage to reduce the computational cost of the subsequent steps.The experimental results show that the proposed method improved the robustness of MSER and achieved 91.4%recall.2.Component Classification with the Strong Classification and Retrieve StrategyThe strong classification and retrieve strategy proposed in this paper includes two stages:strong classification stage and recovery stage.Because the text has rich gradient feature,so in the strong classification stage,we using two adaptive histogram of oriented gradient(HOG)+support vector machine(SVM)constitute the classifier classify the components in different feature spaces.As the strong classification stage will inevitably lose the text component,so in the recovery phase is to get back the text component.In general,the color characteristics of adjacent text components have a large similarity,hence the text components can be retrieved by calculating the text components and non-text components of the color feature similarity.The experimental results show that our method in the text/non-text classification tasks has achieved 94.34%precision.3.CPU-GPU Heterogeneous Parallel SchemeWith the population of multi-core CPU and the development of GPU programming interface,it has been possible to perform small scale parallel calculation on traditional PC.There is a large amount of parallel computing in the text localization algorithm this paper proposed.So a heterogeneous parallel acceleration scheme is presented,which makes full use of the computing resources of CPU and GPU by task partitioning,to speed up the algorithm.The experimental results show that the parallel scheme achieves 12.5x speedup.In this paper,two effective solutions are proposed for two core problems in Uyghur text localization in complex background images.Our method has achieved 81.4%recall and 94.89%precision on the Uyghur in complex background image dataset(UICBI400),and reached the state-of-the-art.Furthermore,we present an effective heterogeneous parallelization scheme for the proposed method,which reduced the time cost.