Find in Library

Abstracts: In the art design and visual effects industry, efficient rendering technology is crucial for enhancing the visual experience and meeting user aesthetic needs. However, the traditional Monte Carlo path-tracing rendering method has the problems of high computational cost and more noise, so the study proposes a network structure combining direct prediction and kernel prediction to construct a denoised Monte Carlo rendering model on the basis of the traditional interactive frame-rate neural network structure and verifies its effectiveness. The experimental results showed that, while the other scenarios yielded the same results, the bedroom scenario's root-mean-square error value was 5.83 10–3, its peak signal-to-noise ratio was 29.34, and its structural similarity value was 0.95, all of which were superior to the comparison method. In the model ablation experiments, the average peak signal-to-noise ratio and structural similarity values were higher than those obtained using the comparison selection method. The average root mean square error of the joint prediction network constructed in the order of direct prediction network and kernel prediction network was 8.39 10–3, which was lower than the comparison selection method. The average peak signal-to-noise ratio when utilizing the adaptive time accumulation module was 31.07, however the maximum peak signal-to-noise ratio when employing bilateral filtering to create kernel weights was up to 35.35. When compared to the standard denoising Monte Carlo rendering model, the proposed model performed strongly better in model comparison experiments and outperformed it in model ablation experiments, providing guidelines for enhancing the quality of images used in art design applications. This study not only proposes innovative solutions to technological challenges, but also provides a theoretical and experimental basis for improving image quality in real-time art and design applications.