Facial Image Makeup Migration via Feature Region Stitching

The patent divides the face image into multiple face parts (eyes, eyebrows, nose, mouth, cheeks) and processes each part separately through dedicated sub-networks. This segmentation approach allows the system to handle pose and shape variations more effectively by focusing on local features rather than treating the entire face as a single region, thereby improving segmentation precision while maintaining processing efficiency.

The patent introduces a multi-dimensional feature extraction approach by generating both original feature maps from the target face image and reference feature maps from the source face image. These feature maps are processed through multiple convolutional layers to extract features at different levels of abstraction, enabling the system to capture both local makeup details and global face structure, thus improving segmentation precision without significantly increasing processing time.

The patent introduces feature maps as an intermediary representation between the segmented makeup regions and the final aligned output. By generating original feature maps that capture the target face's geometry and reference feature maps that capture the source makeup characteristics, the system can perform alignment in the feature space rather than directly in image space. This intermediary approach enables more precise alignment that adapts to pose and shape variations while maintaining implementation feasibility through standard convolutional operations.

The patent performs makeup migration calculations that dynamically adjust alignment parameters based on the extracted features. The system calculates transformed feature maps by combining original and reference feature maps with learned transformation parameters, allowing the alignment to adapt to different poses and shapes. This parameter adjustment mechanism improves alignment precision by capturing subtle geometric variations that fixed grid alignment cannot handle.

The patent applies different processing strategies to different face parts based on their local characteristics. Each face part (eyes, eyebrows, nose, mouth, cheeks) has its own dedicated sub-network that extracts features specific to that region. This local quality approach ensures that makeup migration on each face part considers the local pose and shape characteristics, resulting in more natural fusion effects while maintaining overall processing efficiency through parallel processing of multiple face parts.

The patent implements a dynamic feature extraction and alignment process that adapts to pose and shape variations in the input images. The convolutional networks learn to dynamically adjust feature representations based on the input data, and the alignment process dynamically transforms features to account for pose and shape differences. This dynamic approach improves fusion naturalness by adapting to varying conditions rather than applying fixed processing rules.