HTD

HTD decouples classification and regression tasks in two-stage object detectors to mitigate task-misalignment and improve detection accuracy and localization.

Key Features:

• Task-Specific Decoupling: HTD decouples the sibling heads for classification and regression, enabling task-tailored feature processing to mitigate semantic inconsistency and task-misalignment.
• Progressive Graph (PGraph) Module: Employs progressive graph reasoning with local spatial aggregation and global semantic interaction to enhance semantic representations of region proposals and improve classification.
• Border-Aware Adaptation (BA) Module: Integrates multi-level features adaptively with emphasis on low-level border activation to enhance spatial and border perception for regression-based localization.
• Semantic Feature Aggregation (SFA) Module: Aggregates global semantics using image-level supervision to provide shared semantic knowledge across decoupled branches.
• Instance-Level Semantic Consistency (ISC): Maintains instance-level semantic consistency by utilizing aggregated knowledge from the SFA module to prevent inconsistencies between branches.

Scientific Applications:

• Medical imaging: Improves accuracy and consistency of two-stage detectors for object identification and classification in medical images.
• Biological data analysis: Enhances detection and classification of objects in biological datasets requiring precise localization and semantic discrimination.
• Automated microscopy: Facilitates reliable object detection in microscopy images by improving classification and border-aware localization.

Methodology:

Systematic decoupling of tasks with specialized modules including semantic aggregation, progressive graph reasoning (local spatial aggregation and global semantic interaction), and adaptive multi-level feature integration focused on low-level border activation.