Research And Application Of Natural Scene Text Detection Algorithm

Text detection of natural scene image is a key step for machine to understand the images,It has important practical significance and application value,can be widely used in image text retrieval,machine translation,blindness support systems and so on.Natural scene images are easily affected by noise,illumination and camera angle.The font,shape and size of the text are different,and are mixed with various languages.These mentioned above cause images complex,to be blurred and in low contrast,which increases the difficulty of text detection.Therefore,it is necessary to design a text detection algorithm with strong practicality and robustness.With the development of economy and the increase of currency circulation,Anti-counterfeiting function of currency urgently needs to be enhanced.As one of the anti-counterfeiting technologies,the accuracy of banknote serial character identification can directly affect the currency circulation security.In the field of location and recognition of banknote serial character,the background of RMB image is complex.The RMB images collected by the currency detector are easily affected by light,noise and delivery angles,which increase the difficulty of identifying the banknote serial character.The natural scene text detection algorithm can be effectively applied to the detection of banknote serial character.The main content of the full text is as follows:(1)In order to solve the problem of the image in low contrast and easily affected by noise,the paper utilizes directional gradient information to enhance the contrast of background and text,which enhance the performance of traditional MSER to extract character candidates.Because the human eye is sensitive to color and contrast information,the paper chooses the combination of MSER and clustering in Lab color space to extract character candidates.The algorithm is tested in ICDAR 2003 and ICDAR 2013,f-measure are 0.74 and 0.79 respectively.The algorithm effectively improves the problem that low-contrast images and noise-affected images can not be extracted text correctly.However,the algorithm is not robust to images with complex background,blurred fonts and so on.(2)In order to solve the problem of the image with complex backgrounds,blurred fonts,different shapes and affected by objects similar to text easily,the paper implements a text detection algorithm based on deep learning.The text lines consist of a series of strong correlation characters,these characters have big difference,when text lines are used as a whole object to detect texts,the problem of incomplete detection is likely to occur.Therefore,the paper utilizes Long Short-Term Memory(LSTM)network to strengthen the memory function of network in order to extract more character sequences.Besides,Multi-directional anchors are used for detecting multi-direction text.The algorithm is tested in public datasets ICDAR 2003,ICDAR 2013 and ICDAR 2015,f-measure are 0.84,0.8853 and 0.80 respectively.Experimental results show that LSTM-based text detection algorithm effectively improves the text detection rate.(3)In order to solve the problem of the traditional banknote serial character detection system relies on the spatial location,the algorithm of scene text detection based on LSTM is applied to the banknote serial character detection system.The algorithm solves the problem that the image with angle,stretch and deformation can not be detected characters correctly.The algorithm is tested in the images collected by CIS and achieves 99%detection accuracy.Experimental results prove that LSTM-based text detection algorithm has strong applicability and stability.