| Ceramic tile has the characteristic such as design beautiful,waterproof wear-resisting,it is the indispensable material in decorating design.Style of ceramic tile design automation degree is low,the designer primarily through scanning natural marble or wood grain in designing aspects such as texture,the texture design style style usually single,unable to meet the demand of diversified design,and the traditional method based on image texture generated in ceramic tile realistic and diversification.In order to solve this problem,this thesis proposes a tile image diversity design method based on deep learning,that is,two methods of tile texture automatic generation and tile style transformation.These methods can meet the needs of tile image texture and style diversification.On this basis,this thesis developed a ceramic tile image design system,to help designers convenient and efficient ceramic tile image design.The main work and innovations of this thesis are as follows:(1)To meet the requirement of tile texture diversity,this thesis proposes an automatic generation network of tile image based on sketch matching.Users only need to input one or more hand-drawn sketches in the network to generate a series of diversified tile images that conform to the texture distribution of the sketches.Firstly,a self-built tile data set is used to pre-train the tile generation network.Then,the quality of sketch generation is improved by optimizing the generator structure of sketch generation network.Then,according to the texture distribution characteristics of tile images,a new input sketch style is proposed and two antagonistic loss functions are introduced to guide the tile generator to fit in the direction of the sketch style,which improves the convergence speed of the network and the quality of the generated tile image,and realizes the output of the specified style image.Finally,this thesis proposes a method of implicit space disturbance and interpolation for tile image generator to realize tile series design,that is,to meet the design requirements of multiple series patterns in production.(2)In view of the needs of tile style diversity,this thesis proposes two tile image generation algorithms based on style transfer.One is the multi-stage local style coordination algorithm of tile image,which can coordinate the style between tile design large picture and non-tile texture pattern.First of all,the shallow fusion of tile texture and paste material is realized through rough local style transfer.Then,through delicate local style coordination,the space outliers caused by independent style mapping in the previous step are eliminated,and the depth coordination of tile style and paste pattern is realized.At the same time,according to the characteristics of tile texture distribution,the style content loss function is added to further improve the visual quality of the image.Finally,a block matching image post-processing algorithm is used to eliminate the discordant artifacts,so as to obtain a higher quality tile generated image.Compared with the traditional Poisson image fusion algorithm and the classical style transfer algorithm,this algorithm has better effect.The other is the global style transfer algorithm of tile image.In this thesis,the attention mechanism module is added on the basis of the fast style transfer network,and the block matching post-processing operation is carried out on the generated stylized tile image,which improves the image quality and style transfer effect of the generated tile.(3)In view of the diversified requirements of tile image design,this thesis integrates the proposed algorithm into the system and develops the tile image design system,realizing the function interface of automatic generation of tile image and diversified conversion of tile image style.The main functional modules include sketching tile design module,tile texture diversification module,tile local style coordination module,tile global style transfer module and other modules,providing a convenient and efficient design system for tile image designers. |