Monocular Depth Estimation in Degraded Visual Environments: A Dynamic Context Modeling Framework with Differential Feature Prioritization

Self-supervised monocular depth estimation encounters substantial challenges in scenarios with low-resolution and motion-blurred images. In such scenarios, traditional methods have difficulties dealing with noisy artifacts and the degradation of feature extraction. To tackle these limitations, this paper presents a lightweight framework that combines global context modeling with efficient spatial-channel feature fusion. The proposed approach utilizes a state-space architecture augmented with self-attention mechanisms to capture the long-range dependencies within degraded inputs, overcoming the issues of insufficient feature extraction for low-resolution and motion-blurred images as well as the interference caused by noise. By integrating parameterized differential attention to balance the interactions between global and local features and a dual-path spatiotemporal aggregation module for multi-scale adaptation, this method optimizes computational efficiency while ensuring robust depth prediction accuracy.Experimental validation on benchmarks shows that the framework is capable of reducing hardware dependency and achieving performance on par with that of models trained on high-quality images, thus promoting practical deployment in resource-constrained environments.