| As high-level semantic information,text is an important carrier of human thoughts and feelings.It contains very valuable information and is indispensable to people's daily life.With the development of the Internet and the popularity of mobile terminals such as smartphones and digital cameras,a large number of images have emerged,especially natural scene images,text information in which is not only an important supplement to the scene,but also a very important clue for scene understanding.Therefore,text detection in natural scene images has become one of the research hotspots in recent years.It is widely used,such as human-computer interaction,image search,industrial automation and license plate recognition.For the traditional optical character recognition technology,there have been quite mature solutions,which has made remarkable achievements when dealing with document texts.However,due to the diversity of text,the complexity of background and the interference of other external factors,text detection in natural scene images still faces many challenges.In order to solve the problem of low precision in existing natural scene text detection algorithm,a text detection algorithm based on improved convolutional neural network and the feature of text lines was proposed,which improves the algorithm from three aspects: extraction and pruning of connected components,classification of connected regions,formation and classification of multidirection candidate text lines.The main contributions of this thesis are as follows:(1)The enhanced maximally stable extreme region was used to extract connected components of the image which can segment the character pixels linked together due to ambiguity and the holes in the character connected regions.The nested maximally stable extreme region was pruned by pruning operation combined with smoothness to obtain isolated connected regions,which is convenient for subsequent classification of connected components.(2)When classifying connected regions,the traditional convolutional neural network algorithm was improved.In order to balance the relationship between precision,recall,time complexity,the number of convolution levels and threshold,the optimal number of convolution layers and threshold are determined through a large number of experiments.The proposed algorithm used four-level convolution which has stronger ability to learn features,which is enough to extract deep features of characters and background,and can significantly improve the precision of text detection and background and does not increase too much computation at the same time.The proposed algorithm reduced the threshold,which greatly improves the recall of text detection,and at the same time,the precision will not be reduced too much,because four-level convolution increases the precision which is enough to compensate for the decline in precision at this time.(3)In the process of the formation of multi-oriented candidate text lines,a character merging method based on the feature of text lines was proposed,which is easy to operate and has amazing effects.In the process of multi-oriented candidate text line classification,a classification algorithm based on C4.5 decision tree was proposed.Machine learning algorithm was used to further classify candidate text lines to get the final text line because of its robustness.This thesis improves the text detection algorithm from the above three aspects.Experiments were carried out on ICDAR2013,ICDAR2015 and MSRA-TD500 datasets respectively.Experimental results show that the proposed text detection algorithm can achieve good detection performance in dealing with multi-oriented text images which are affected by negative factors such as blur,perspective distortion and extreme illumination.The proposed algorithm can significantly improve the precision and recall of text detection in natural scenes,and is suitable for text in any direction,language and font.Therefore,the proposed algorithm has better text detection performance and higher robustness. |
PDF Full Text Request