CMDSR

CMDSR implements a conditional meta-network for single-image super-resolution (SISR) to adapt SR models to multiple simultaneous degradations and improve reconstruction across varying degradation distributions.

Key Features:

• Conditional Hyper-Network Framework: CMDSR employs a conditional meta-network that dynamically adjusts super-resolution model parameters according to input degradation characteristics.
• ConditionNet and BaseNet Integration: ConditionNet extracts task-level degradation priors from a support set of degraded image patches which are used to adaptively modify BaseNet parameters.
• Task Contrastive Loss: Training uses a task contrastive loss that minimizes inner-task feature distances and maximizes cross-task feature distances to separate degradation priors.
• Blind Framework with Single Parameter Update: CMDSR operates as a blind framework that does not require predefined degradation maps and performs adaptation with a single parameter update.
• General Applicability: The flexible BaseNet structure enables application across a range of SISR models and degradation scenarios.

Scientific Applications:

• SISR under multiple degradations: Adaptation of super-resolution models to inputs affected by multiple simultaneous degradations.
• Medical Imaging: Improve reconstruction quality in medical images subject to complex degradations.
• Remote Sensing: Enhance resolution and reconstruction of remotely sensed imagery with heterogeneous degradations.
• Digital Forensics: Support image enhancement for forensic analysis where images suffer varied degradation patterns.

Methodology:

CMDSR uses a conditional meta-network where ConditionNet extracts task-level degradation priors from a support set of degraded image patches to modulate BaseNet parameters, and training employs a task contrastive loss (minimizing inner-task and maximizing cross-task distances); adaptation is performed blind via a single parameter update.