Print Character Recognition Based On Wavelet Analysis And CNN

With the rapid economic development,digital transformation has become a future trend.As a key technology to realize digitalization,character recognition technology has always been a hot research topic of people's attention.Character recognition can be widely used in the fields of transportation,finance,industry,economy and artificial intelligence.In-depth study of character recognition technology has good practical significance for promoting the development of technology and economy.In this context,this article takes the US dollar banknotes in the field of financial circulation as an example.Taking into account the uniqueness of the US dollar serial number,it is automatically identified and used as an identifier in circulation,which can effectively monitor its circulation trajectory and also hit the economy.Crimes,etc.;therefore,combined with wavelet theory and convolutional neural networks,the method of extracting and identifying US dollar prefix numbers and their implementation methods are studied in detail.The specific work is as follows:(1)A US dollar number image preprocessing method based on the combination of bilinear interpolation and morphological reconstruction image enhancement is given.Because the image of the US dollar number that was collected affects the character definition due to noise and other factors,the paper first uses bilinear interpolation to enlarge the image and realizes the image rotation according to the detected angle,and then reconstructs the image according to the morphology that can effectively distinguish image characters with different light intensities and the background,highlight the characteristics of the characters to realize image enhancement,and get the preprocessed dollar number image.(2)The character extraction method based on wavelet analysis,morphology,and projection method is given.In order to better realize character recognition,this paper uses wavelet analysis combined with conventional image segmentation methods to realize character positioning.Because the multi-scale characteristics of wavelet analysis can overcome the interference of noise in edge detection,and can avoid the loss of details caused by the picture in edge detection,the paper first uses wavelet analysis to perform edge detection on the preprocessed dollar number image;Value and morphological hole filling are used to locate image characters;finally,three projections and character normalization are used to achieve character extraction.(3)A character recognition algorithm based on improved CNN is given.Because the traditional convolution neural network will misjudge similar characters when recognizing,the paper has put forward an improved CNN character recognition algorithm after many researches and tests.First,build a first-level CNN network model,change the size of the input layer and convolution kernel based on the structure of the traditional CNN network model,improve the recognition speed of characters,and achieve "coarse division" of characters;secondly,build a second-level CNN network model,improve the series of the traditional CNN network model,perform secondary recognition processing on easily confused characters to improve the accuracy of character recognition,and achieve character "subdivision".Finally,combined with the above-mentioned series of methods,the actual collected US dollar number image is recognized on the DSP(DM642)chip,which achieves the expected effect and verifies the feasibility and superiority of the algorithm.