Abstract
Techniques are disclosed for multi-stage image segmentation with selective high-resolution region refinement. A coarse segmentation neural network generates a full-frame mask while operating with a reduced internal resolution to meet real-time constraints. From coarse outputs, a region of interest (ROI), such as a head/hair region, is dynamically determined per image or per video frame. The ROI is cropped from a higher-resolution version of the original input image, optionally along with aligned coarse segmentation data, and provided to a second refinement neural network that produces refined alpha or mask values and confidence. A fusion stage combines refined ROI output with the coarse full-frame output using confidence-gated blending and a spatially varying boundary transition to avoid seams, and falls back to coarse results when refinement confidence is low. Implementations include single-graph networks or two-model pipelines with GPU-based cropping and fusion.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Multi-Resolution Image Segmentation with Selective Region-Based Detail Refinement", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10623