Weakly Supervised Multi-Source Fine-Grained Image Recognition via Class-Conditional Adversarial Adaptation and Confidence-Gated Pseudo-Labeling

Abstract

Differences in devices and imaging styles across multi-source images lead to distribution shifts between training and deployment. Under weak annotation, models supervised only by image-level labels are easily affected by background textures, causing category confusion and pseudo-label noise accumulation in target domains. To address this issue, this paper proposes CCAP-Net, a weakly annotated multi-source subcategory recognition network based on class-conditional adversarial adaptation and collaborative confidence pseudo-labelling. CCAP-Net introduces complementary component attention branches to mine diverse local discriminative regions and fuses global and local representations through a gating mechanism to enhance subtle subcategory cues. A conditional domain discriminator is further designed to align domain distributions while preserving class-related structure. To improve robustness under multi-source imbalance and unreliable target predictions, adaptive source contribution learning and confidence gating are incorporated to dynamically regulate source influence and suppress low-confidence target samples. In addition, a teacher network generates pseudo labels for unlabeled target data, and target training combines strong/weak augmentation consistency with class-centre similarity filtering to reduce error propagation. Experiments on Office-Home, DomainNet, and CompCars demonstrate that CCAP-Net consistently improves target-domain accuracy, class-balanced metrics, and calibration reliability over representative baselines.