Descrição do Prompt

Create a clean, professional framework diagram for a computer vision research paper submitted to ICMR. The diagram illustrates a dual-branch semi-supervised learning framework for realistic long-tailed data. Structure: On the left, show two data streams: Labeled Data (Long-Tailed) and Unlabeled Data (Unknown Distribution). Both streams pass through a shared feature extractor (Backbone Network). After feature extraction, the features split into two parallel branches: Biased Predictor: Generates pseudo-labels for unlabeled data Annotated as “adaptive to unlabeled data distribution” Balanced Predictor: Trained to encourage balanced decision boundaries Annotated as “used for final inference” Between the backbone and both predictors, include a module named Learning-Status-Guided Feature Diffusion Suppression, connected to the feature space. Show arrows indicating suppression of feature expansion for well-learned classes and preserved exploration for under-learned classes. Pseudo-labels generated by the Biased Predictor are fed back to supervise training. Style requirements: Minimalist academic style Flat vector graphics Soft blue and gray color palette Clear arrows and labeled components No clutter, no decorative elements White background Suitable for conference papers (ICMR / CVPR / ICCV style) Text labels should be clear and readable, with consistent font size and alignment.