Research On Grayscale Image Coloring Of Ethnic Costumes Based On Deep Learning

Using digital technology to overcome the limitations of black-and-white forms of ethnic culture image resources left over from the previous century and current ethnic costume workmanship has become the favored way.Deep learning’s advancement gives a new theoretical foundation for grayscale picture coloring challenges.Deep learning can construct mapping associations directly,eliminating artificial influence,and coloring quality is dependent on model correctness,making completely automated coloring feasible.Traditional coloring techniques take a great deal of human direction and provide low-quality coloring.However,the majority of existing deep learning coloring algorithms are based on low-resolution research,which seems inadequate for the job of coloring particular scenes.Ethnic costumes are characterized by rich semantic information,numerous stripes and veins,diverse colors,and high resolution,whereas mainstream methods are primarily applicable to low-resolution images,with the entire image as the coloring object and insufficient attention to the detailed texture part,these factors make it challenging to demonstrate good results on the dataset,and there are issues with inaccurate coloring or even distortion in some regions.Therefore,it is difficult to color high-resolution photos of ethnic clothing classes in grayscale.In this study,we enhance the generative adversarial network from a practical standpoint and evaluate and compare the grayscale picture coloring approaches suited for ethnic clothing styles.This paper’s primary body consists of the following four sections:(1)In the topic of ethnic clothes,a more comprehensive photographic database has been built.More than two thousand high-definition costume images of four ethnic minorities in Yunnan were collected by crawling from ethnic websites and visiting ethnic gathering areas and museums in the field.The samples with poor quality were screened out through post-processing and the size of the dataset for future research was adjusted accordingly.(2)An adversarial coloring approach for ethnic costume photos that combines global characteristics and an attention strategy is designed.On the basis of the end-to-end network architecture,global features are extracted and injected into the generator as a mask;the attention mechanism is used to constrain the feature changes in the network to improve the training stability;and a multi-scale discriminator is used to capture the subtle differences in scale between the generated samples and the real samples to force the data generation distribution to converge to the real distribution.Multiple sets of comparison studies demonstrate that the model performs better on photos of ethnic costumes.(3)A grayscale image coloring method of ethnic costumes based on the fusion of style and content features is designed.To address the difference in style between the coloring images and the real images,in this paper,we adjust the network structure and training strategy appropriately within the Style GAN framework,adopt dynamic convolution to improve the focus on the convolution kernel and enhance the model’s performance laterally,and combine the multi-scale upsampling module to extract deeper features from the output results at multiple scales for the upsampling process.The benefits and drawbacks of the model and the prevalent approaches are examined in a variety of ways,and it is shown that the method presented in this study is closer to the actual scenario in terms of coloring image styles and has applicability in the area of ethnic costumes.(4)The interactive ethnic costume coloring system is designed and implemented.The system is implemented using Tensorflow framework and Python-related tool library,using C/S architecture,providing core functions such as image processing and image coloring for users,collecting feedback information to continuously improve the system and spreading the culture of ethnic costumes.