Research On The Deep Generation Algorithm And Its Applications In Style Transfer Of Calligraphy And Painting

Due to the development of deep learning,image processing techniques have achieved revolutionary progress,especially in the field of classification and identification where their performances are already approaching or have even surpassed that of human beings.However,these achievements are mainly successes in the realm of pattern recognition,which means there is still a long way to go before the arrival of true artificial intelligence.Among those techniques,deep generation algorithm has found a wide range of application prospects due to their powerful modelling capabilities,which justifies the choice of the core content in this thesis,the study of generative models based on deep learning.Two applications of generation models,transfer of painting one-shot learning for painting style transfer and multi-sample learning for calligraphic stylization learning,were taken as the drive in this thesis.Improvements in networks based on generator-discriminator models and discriminator models were proposed,enhancing the performance of the original neural style transfer algorithm and realizing the calligraphic stylization learning.In particular,the imitation of famous calligraphist's cursive handwriting was achieved.The detailed research contents and innovations of this thesis are as follows.(1)One-shot learning for painting style transfer.A zigzag learning method to enhance the effects of style learning was proposed,and the architecture of zigzag learning to multi-sample and cost functions were expounded in detail.Compared to the original painting transfer algorithm,even under the case when the contents of the style image and the content image were distinctively different,the improved method managed to reduce the reconstruction deviation in structural information while mitigating the artefact effect for areas with low texture or no texture in the generated images.Finally,an evaluation method to quantitatively characterize the performance of the proposed algorithms.(2)Calligraphic stylization learning in terms of multi-sample learning.In this chapter a novel active generation networks in analogy to the active shape model and its training method were proposed that ameliorated the difficulty in convergence and training for generative adversarial networks.The framework was applied in calligraphic stylization learning.Combined with embedding layer,we exploit information across multiple fonts,thus enhancing the generalization ability of the model and realizing simultaneous learning of several styles in the network.Multiple regularization approaches were performed to the cost function by introducing the multipath discriminative network,thereby realizing transfer result with a pixel level precision.(3)Incremental learning of discriminator based on neuron differentiation.When the generative adversarial network learns in a new domain,discriminator in the calligraphic stylization learning framework needs training all over again,which is rather time-consuming.In this chapter,an incremental learning method based on neuron differentiation was proposed in the discriminator in which a “clustering layer” and a “differentiation layer” were added to perform incremental learning to the input category in the form of self-organization.According to validation results in public classification datasets,the proposed method had a higher accuracy in classification when the network size was equal.Also,the learning performance of the proposed algorithm was tested in new fonts,manifesting fewer time costs under the same condition of convergence.