Research On Calligraphy Font And Character Synchronous Recognition Based On Deep Transfer Learning

The number of Chinese calligraphy characters is huge,and different fonts have different unique features.In the early years,manual calligraphy character recognition required people with professional training,which brought difficulties in the popularization of calligraphy culture.With the development of artificial intelligence technology,people have begun to use machine learning algorithms to realize automatic calligraphy character recognition.Chinese character font recognition(CCFR)and Chinese character recognition(CCR)technologies have been widely used in various scenes.However,in some specific scenes(such as calligraphy teaching and calligraphy evaluation),both calligraphy font information and character information are required.And there is no research on synchronous recognition of calligraphy font and character by combining calligraphy font recognition task and calligraphy character recognition task.Therefore,this thesis proposed a method for synchronous recognition of calligraphy font and character based on deep transfer learning.First of all,since there is no complete dataset suitable for this synchronous recognition task,a special dataset for synchronous recognition of calligraphy font and character was created by using image data enhancement technology: We used a font file to generate a basic training dataset and then used grayscale transformation,inverse transformation,dilation corrosion transformation,perspective transformation,adding speckle noise and other methods to expand the dataset.Finally,a calligraphy recognition training set containing 80,000 calligraphy word images was constructed.Secondly,a deep convolutional neural network for font and character synchronous recognition(FACSR-CNN)was built,which trained on the specific dataset for calligraphy recognition.Based on the shallow feature sharing scheme in deep transfer learning,a calligraphy font recognition network and 5 calligraphy character recognition networks were combined to achieve the goal of multi-task information sharing,which also greatly reduced computing consumption.On this basis,multiple networks' output probability product(MNOPP)algorithm was proposed for the synchronous recognition of calligraphy font and character.Through multiplying the output probability sequence of calligraphy font recognition and the output probability sequence of calligraphy character recognition,the output information can be well integrated.Therefore,the two recognition tasks can be corrected mutually,and the accuracy of synchronous recognition was significantly improved.Finally,experiments show that the method in this thesis achieves 95.6% synchronous recognition accuracy on the test dataset.As the first synchronous recognition method of calligraphy font and character based on deep transfer learning,it will have broad application prospects.