Human-Machine Collaborative Model for Dynamic Translation of Intangible Cultural Heritage Patterns Based on Generative Adversarial Network and Reinforcement Learning from Human Feedback

At present, there is a lack of cultural and artistic expression in the modern design transformation of Intangible Cultural Heritage (ICH) patterns. For this reason, this study proposes a dynamic translation human-machine collaborative model that combines the Generative Adversarial Network (GAN) with Feedback Learning from Human Feedback (RLHF), with the aim of improving the cultural accuracy and quality of the generation of ICH patterns by means of human-machine collaborative. The model includes the content and style extraction module based on conditional Generative Adversarial Network (cGAN) and RLHF mechanism. The former is used to separate and extract the structure and aesthetic features of patterns, while the latter uses reward model to learn expert preferences, and selects near-end strategy to optimize Proximal Policy Optimization (PPO) algorithm to iteratively optimize the generation process. The model is verified on a dataset containing a large number of embroidery images, and it is found that the objective index of the model is better than that of the contrast model, with Fréchet Incidence Distance (FID) of 28.5 and Learned Perceptual Image Patch Similarity (LPIPS) of 0.162. In the subjective evaluation, the score of cultural fit is 4.35 (5-point scale), the designer's task completion time is shortened, and the cognitive load is reduced from 6.2 to 3.1. The above results show that the proposed model can effectively transform the non-legacy patterns with high quality, and strengthen the content style while maintaining the basic cultural characteristics of the patterns, which provides certain technical ideas for the cultural inheritance and innovation of man-machine collaboration.