Study On Text Detection In Natural Scene Images

With the popularity of mobile terminals,such as mobile phones and panels,natural scene images captured by mobile devices prompt the emergence.Text,as one of the most influential inventions of mankind,plays an important role in human production and life,which is one of the important components of natural scene images.It is rich in semantic information,and it is an important carrier of human thought and emotion expression,which can be seen everywhere in natural scenes.People tend to pay more attention to the natural scene text content compared with other objects.Text detection and recognition is of great significance to scene image analysis.This thesis describes the current difficulties and challenges in the field of text detection in natural scene images in detail,as well as common text detection features and methods.In addition,it analyses the advantages and disadvantages.After that,a novel text detection algorithm based on single exemplar is proposed.By calculating the feature similarity between single exemplar and target image,the text regions in the natural scene image is initially located,which can greatly reduce the non-text regions proposed in the MSER algorithm.For the text candidates,some geometric constraints and the stroke width feature are used to remove non-text regions.Finally by combining text blocks with similar features,the text areas are extracted.The algorithm addresses the problems of the learning algorithm which requires a large number of training samples,and slow training speed.In addition,it also addresses the problem that MSER algorithm extracts a large number of non-text regions for scene images with complex background.The experimental results show that,the proposed algorithm can accurately detect the text areas in the image to achieve desired results.This paper also proposes a learning based text detection algorithm using multi-channel MSER,which effectively tempts the problems that traditional MSER algorithm is sensitive to fuzzy,low contrast,and uneven illumination.First,the MSER regions are extracted as the character candidate regions for the color edge image and the contrast enhancement image.Then,an ideal text and non-text classifier is trained by SVM,combining with the high level features extracted by convolution neural network,which enhances the performance of text detection algorithm.