A Gradient-Decoupled Multi-Task Learning Framework for Safety-Oriented Object Detection in Degraded Visibility

Visual degradation under degraded visibility distorts low-level feature representations, leading to critical safety risks in industrial monitoring systems. While conventional approaches either incur substantial computational overhead or suffer from training instability, this paper proposes a Gradient-Decoupled Multi-Task Learning Framework that emphasizes feature-level regularization rather than direct image restoration. To mitigate task interference, we introduce Selective Gradient Blocking, which detaches restoration gradients from P4 and P5—the earliest layers of semantic abstraction—to preserve detection-critical representations. Evaluation is conducted under an intentional stress-test configuration using synthetically degraded visibility to establish a robustness floor for safety-critical detection. Experimental results demonstrate a substantial reduction in Head-level false negatives and an improvement in recall, alongside improved stability of prediction confidence. Beyond aggregate accuracy, the proposed framework proves effective as a generalizable training principle for stabilizing detection under degraded visibility.