Research On Tile Texture Image Generation Technology Based On Deep Learning

With the improvement of home living environment,people’s demand for high-quality ceramic tiles is becoming increasingly urgent,leading to a higher requirement for personalized tile image design.At present,tile image design usually adopts the traditional method of texture scanning and post-editing,which has a low degree of automation and a long design cycle.It is difficult to meet a large number of personalized design requirements in a short time.With the improvement of deep learning,techniques such as content generation and style transfer have been proposed to facilitate tile image design work,but in practical applications,there are still limitations in terms of precision,style diversity and texture controllability of generated images,and lacking of related improvement research.In this thesis,we propose three deep learning-based algorithms for personalized generation of tile texture images,to bring assistance and inspiration to the design process and help designers to design tile images more conveniently and efficiently.(1)For the demand of high precision in tile images,this thesis proposes a multi-stage tile texture image generation method based on Generative Adversarial Network(GAN).The method divides the tile image generation process into three stages: tile texture image generation,tile texture image stylization,and tile texture image magnification,and takes the output results of the former stage as the latter input.First,the multi-scale attention Style GAN is used in the first stage,that is,the multi-scale attention weight connection is added to the Style GAN convolution layers based on the same size discriminator and generator feature map.As a result,the multi-scale texture features can be automatically trained to bring rich texture for generating tile images.Secondly,the GAN with smoothing activation is used in the second stage.The activation function of texture synthesis GAN is smoothed and improved in the stylization process,so that the oscillation of activation values is reduced to generate detail-enhanced tile texture images.Finally,through the style iteration mechanism and the magnification method based on super-resolution network,the final tile texture images can be automatically generated with larger-size.The experimental results verify that the algorithm can generate large size and high quality tile texture images with faster generation speed compared with traditional methods.(2)For the demand of style diversity in tile images,this thesis proposes a classifier-guided multi-style tile texture image generation algorithm.First,the classifier-guided Style GAN is applied.It takes the classification condition vector as input,and directs the classifier-guided generator module to influence the output,so that the generated tile style images have higher style attributes.Secondly,the original texture image is chunked,then the texture blocks are trained with the tile style images by Ada IN-GAN to form new tile texture image blocks with global stitching.Next,a linear weighting-based seam elimination method is adopted,meaning that the horizontal,vertical,and center seams are filled with linear weighted image block overlays,thus achieving the purpose of smoothing the seams and improving the overall quality of the new tile texture image.Finally,a series of style-fused tile images with smooth transitions are generated by the weighted fusion of different input classification condition vectors.The experimental results demonstrate that the new tile texture image generated by this algorithm has diverse styles and meets the customer’s demand for tile style diversity.(3)For the texture controllability requirement of tile images,this thesis proposes a mask-guided tile texture image generation algorithm.First,a global OTSU tile sketch extraction method is used to generate a mask binary sketch by setting a global OTSU threshold for binary segmentation of input hand-drawn texture images.Then,the Mask-Fine GAN tile image generation network is constructed,which contains the feature encoder and Fine GAN local generator.The network converts the tile image blocks into background and foreground feature maps,as well as the mask binary sketch into contour feature map.The contour feature map is used as a mask guide to fuse the background and foreground feature maps to generate tile texture images with clear texture structure.The experimental results show that the mask binary sketch converted from the input hand-drawn texture has an obvious guiding effect on the texture structure of the output tile texture image,which satisfies the design intention of texture controllability.