Character generation and visual quality enhancement in animated films using deep learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Improving animated film visual quality is crucial for effective communication.
• Current methods struggle with accuracy and image distortion, especially with complex backgrounds and pose changes.

Purpose of the Study:

• To optimize the First Order Motion Model (FOMM) for generating high-quality animated character images.
• To enhance accuracy and reduce distortion in generated images, particularly in challenging scenarios.

Main Methods:

• Integration of a redesigned Convolutional Block Attention Module (CBAM) into FOMM to focus on critical features.
• Introduction of a repainting image repair module with multi-scale upsampling and occlusion map prediction.
• Development of an enhanced FOMM (E-FOOM) model coupling attention and reconstruction for robust end-to-end generation.

Main Results:

• E-FOOM demonstrated superior performance over existing models in generated image quality, keypoint detection, and pose reconstruction.
• Significant improvements observed: minimum 1.11 dB increase in Peak Signal-to-Noise Ratio (PSNR) and 0.014 in Structural Similarity Index (SSIM).
• Generated images exhibit enhanced pixel-level, structural, and perceptual quality.

Conclusions:

• The E-FOOM model offers a robust framework for high-quality animated character image generation.
• This work provides a technical pathway for achieving superior visual effects in animated films.
• The integration of attention mechanisms and advanced reconstruction techniques addresses key limitations in current models.