A Controllable Style Transfer Algorithm Based On Dilated Convolution Strokes

Using convolutional neural network to render original image with the style of an artistic painting is a hot topic in academia and industry.Compared with the traditional style transfer algorithm,style transfer based on the convolution neural network has the advantages of speed and quality.However,this method is still immature in controlling the stylization effect of the image.For example,the control of stroke size is not flexible enough,causing the huge difference between the generated texture strokes and the original style image.There have been a lot of researches about this problem.The existing methods are mainly divided into two kinds:based on control network perception field and based on coarse-to-fine image refinement process,and these two kinds of methods are only applicable to non-high-resolution and high-resolution images respectively.Dilated convolution is proved to be a very efficient way to adjust receptive fields without losing resolution,and has been successfully applied in many computer vision tasks.Therefore,on the basis of existing researches of stroke size controllable style transfer this thesis explored the method of stroke control based on dilated convolution,and realized the stroke controllable style transfer algorithm applicable to both non-highresolution and high-resolution images.In addition,this thesis designed and implemented an application program for stroke controllable style transfer,which extends the practicability of this technology.The research results of this thesis are as follows:1.A new stroke controllable style conversion method is proposed,which can effectively generate stylized images with different scale textures and high quality.The range of stroke size control is adjusted by the expansion rate parameter of expansion convolution.When the texture is too large to be captured by the previous method,our method can capture by increasing the expansion rate.3.A parameter setting strategy of extended convolution is proposed to generate an ideal stroke size without blindly adjusting the parameters or producing unnsatisfactory results.This method can adjust the stroke size by changing the parameters of expansion convolution,which is more convenient and flexible than the previous algorithm.4.The desktop application of style migration is designed and implemented,and the GUI extends the application of style migration technology.