Study Of CNN And The Application In Character Recognition

Character recognition plays an indispensable role in the future production of our life,character recognition has been widely applied to all walks of life.The identification and tracking of series number plays a decisive role in the stable operation of the financial Internet;the recognition of license plate number is essential for vehicle management;the guideposts' recognition for the future identification of unmanned driving has great significance;accurate scanning of tracking number for the logistics industry like a tiger which has got wings,and its like.Because of its unique weight sharing the Convolution Neural Network(CNN)also got the favor of many researchers.By studying the influence of various factors on the identification effect,the paper focus on the following three aspects to improve the efficiency of character recognition.Image denoising.Due to the influence caused by environment and light,it will lead to recognition of character image have different degrees of noise.In order to eliminate the unnecessary fuzzy edge,enhance the features of the image,improve the effect of image feature extraction,the paper selected wavelet transform commonly known as "mathematical microscope" to strengthen the image's features and suppress the unwanted noise.Character segmentation.For the general segmentation,paper use the three projection method for character segmentation,and it achieved good results.As for the contaminated character,there is still not a perfect segmentation algorithm for using.Finally,considering the character segmentation has important significance for the character recognition,and in order to overcome the difficulties that adhesion character can not segment,the method of window moving was proposed,and it achieved considerable results.Network structure.Because there exist some similarity between partial characters,using the typical CNN model can not distinguish the character in a good result.After repeatedly testing,the paper puts forward the two-level CNN model for character recognition,and the first-level's output is 23.In summary,a method is proposed to recognize the character number,it is a more efficient,more high-speed,more precision way than before,and it based on the convolution neural network(CNN)which can directly extract characters from the training sample.Aimed at the recognition of dirtied character image,the method of window moving was proposed.And in order to make the features protruding,the paper using wavelet transform which commonly known as "mathematical microscope" in the image preprocessing stage to eliminate the noise of image.The paper based on the banknotes that is collected through the DM642 chip and the number is 100 thousand.it can effectively reach the robustness of the recognition algorithm under some circumstances such as the damaged,smirched identified object,the recognition accuracy can reach at least 99.99% and the recognition time of thesingle identification image can be controlled within 5ms.