The Research On The Methods Of Natural Scene Text Detection

Text detection in natural scene images is an important yet challenging task for many content-based image applications, due to the significant variations of the appearance of the text itself and the complexity of the scene context.This thesis proposes two effective detection methods for complicated scene text: 1) text detection based on robust Stroke Width Transform (SWT) and Deep Belief Network (DBN) and 2) text detection based on Maximally Stable Extremal Region al-gorithm (MSER) and Convolutional Neural Network (CNN). In the former method, we propose a robust extension of SWT algorithm to generate the SWT image from an input scene image, from which we then extract candidate characters by connected component analysis and exploitation of gradient cues. Next, after filtering out invalid candidates by several heuristic filtering rules, a DBN-based neural network is exploited to ex-tract seed candidate characters, which are then used to localize as many as possible the non-seed candidate characters for higher recall rates. Finally, the resulting candidate characters that have similar attributes are aggregated into text lines as the text detection result. In the latter method, a cascaded CNN model is proposed to accurately localize seed candidate characters from the robust MSERs of the input image with the learned discriminative deep convolutional features. Next, an iterative and incremental grow-ing algorithm is proposed to grow from seed candidate characters to search for other (possibly degraded) text components based on their conformity to the seed characters, which are measured by the similarity metric combining both geometric and appearance aspects of the relations between two neighbouring text components.The effectiveness of the two proposed methods have been verified on several stan-dard public scene text datasets. Experiments show that, since the SWT algorithm tightly relies on the edge detection result, the SWT-DBN method could be insufficient for text in complicated scene context, for which the MSER-CNN method generally attains better and state-of-the-art results (especially on the recall measure). In both methods, the proposed text growing strategy and the deep neural network (e.g. CNN) based candidate character localization scheme have demonstrated their effectiveness.