Research On Rendering Method Of National Costume Sketch Based On Deep Learning

The costume culture of ethnic minorities plays an important role in the traditional culture of the Chinese nation.With the continuous development of deep learning and computer vision,it has become an important research topic to use the deep learning framework to achieve the rendering of ethnic minority clothing sketches.In order to better protect and inherit the ethnic costume culture,this thesis proposes a deep learning-based rendering method of ethnic costume sketches.The main research contents and achievements are as follows.Firstly,this thesis proposes an automatic generation model of ethnic costume sketches based on the Cycle GAN network.It realizes the extraction of the edge and contour features of ethnic minority clothing images.The model takes a 512×1024 highresolution rectangular image of ethnic costumes as input.The Generator of the model comprises an Encoder,a Converter,and a Decoder.The Patch GAN is used as the discriminator.In order to enhance the effect of sketch generation,bilateral filtering and K-means algorithm are used in the preprocessing stage to denoise and color quantify the ethnic clothing images.It generates the problem of messy lines in the sketch.Experiments show that the method proposed in this thesis can effectively achieve the automatic generation of ethnic minority clothing sketches,and the details of clothing information are well preserved.Secondly,this thesis proposes a coloring model of ethnic costume sketches based on Pix2 Pix.Based on the CGAN network,in order to constrain the target image generation process,this thesis inputs the original image as a "label" in the generator.The generator of this model uses the Res Net network,which comprises 9 Resnet Blocks(residual blocks).The residual block can add the information before the convolution operation to the feature information extracted after the two-layer convolution operation on the other side as an output.This ensures that it can also use the previously extracted feature information in later networks.The discriminator is a 70×70 Patch GAN.Through training in the data set,the experiment proves that this method can get a better coloring effect.And the comparison with a variety of coloring methods fully shows that the model has good coloring performance.Finally,this thesis proposes a minority costume rendering platform to realize the application of the minority costume sketch generation algorithm and coloring algorithm.The platform integrates functions such as image preprocessing,sketch generation,and sketch coloring.And it shows the major research results of this thesis.The rendering function of costume ethnic style increases the interest of the system and further improves the practicability of the platform.The realization of the platform also indirectly demonstrates the effectiveness of the method in this thesis.