Deploying high-performance image restoration models on drones is critical for applications like autonomous navigation, surveillance, and environmental monitoring. However, the computational and memory limitations of drones pose significant challenges to utilizing complex image restoration models in real-world scenarios. To address this issue, we propose the Simultaneous Learning Knowledge Distillation (SLKD) framework, specifically designed to compress image restoration models for resource-constrained drones. SLKD introduces a dual-teacher, single-student architecture that integrates two complementary learning strategies: Degradation Removal Learning (DRL) and Image Reconstruction Learning (IRL). In DRL, the student encoder learns to eliminate degradation factors by mimicking Teacher A, which processes degraded images utilizing a BRISQUE-based extractor to capture degradation-sensitive natural scene statistics. Concurrently, in IRL, the student decoder reconstructs clean images by learning from Teacher B, which processes clean images, guided by a PIQE-based extractor that emphasizes the preservation of edge and texture features essential for high-quality reconstruction. This dual-teacher approach enables the student model to learn from both degraded and clean images simultaneously, achieving robust image restoration while significantly reducing computational complexity. Experimental evaluations across five benchmark datasets and three restoration tasks—deraining, deblurring, and dehazing—demonstrate that, compared to the teacher models, the SLKD student models achieve an average reduction of 85.4% in FLOPs and 85.8% in model parameters, with only a slight average decrease of 2.6% in PSNR and 0.9% in SSIM. These results highlight the practicality of integrating SLKD-compressed models into autonomous systems, offering efficient and real-time image restoration for aerial platforms operating in challenging environments.

Image enhancement methods leveraging learning-based approaches have demonstrated impressive results when trained on synthetic degraded-clear image pairs. However, when deployed in real-world scenarios, such models often suffer significant performance degradation due to the inherent domain gap between synthetic and real degradations. To bridge this gap, we propose a novel Two-stage Contrastive Domain Adaptation image Enhancement (TCDAE) framework consisting of two key strategies: (1) Synthetic-to-Real Domain Transfer Learning (S2R-DTL) that effectively translates images from the synthetic degraded domain to the real degraded domain, aligning the domains at the pixel level, and (2) Degraded-to-Clear Domain Transfer Learning (D2C-DTL) that further adapts the enhancement model from the synthetic to the real domain by translating images from the real degraded domain to the real clean domain in both supervised and unsupervised branches. A unique aspect of our approach is the integration of a Domain Noise Contrastive Estimation (DoNCE) loss in both learning strategies. This specialized loss formulation enables TCDAE to robustly translate images across domains, even in scenarios lacking strong positive examples. Consequently, our framework can generate enhanced images with natural, realistic appearances akin to real clear images. Comprehensive experiments on real-world degraded scenes across diverse tasks, including dehazing, deraining, and deblurring, demonstrate the superiority of TCDAE over state-of-the-art methods, achieving improved visual quality, quantitative metrics, and downstream task performance.

In the wild, snow is not merely additive noise; it is a non-stationary, semi-transparent veil whose spatial statistics vary with depth, illumination, and wind. Because conventional two-stage pipelines first detect a binary mask and then inpaint the occluded regions, any early mis-classification is irreversibly baked into the final result, leading to over-smoothed textures or ghosting artifacts. We propose BAT-Net, a Bidirectional Attention Transformer Network that frames desnowing as a coupled representation learning problem, jointly disentangling snow appearance and scene radiance in a single forward pass. Our core contributions are as follows: (1) A novel dual-decoder architecture where a background decoder and a snow decoder are coupled via a Bidirectional Attention Module (BAM). The BAM implements a continuous predict–verify–correct mechanism, allowing the background branch to dynamically accept, reject, or refine the snow branch’s occlusion hypotheses, dramatically reducing error accumulation. (2) A lightweight yet effective multi-scale feature fusion scheme comprising a Scale Conversion Module (SCM) and a Feature Aggregation Module (FAM), enabling the model to handle the large scale variance among snowflakes without a prohibitive computational cost. (3) The introduction of the FallingSnow dataset, curated to eliminate the label noise caused by irremovable ground snow in existing benchmarks, providing a cleaner benchmark for evaluating dynamic snow removal. Extensive experiments on synthetic and real-world datasets demonstrate that BAT-Net sets a new state of the art. It achieves a PSNR of 35.78 dB on the CSD dataset, outperforming the best prior model by 1.37 dB, and also achieves top results on SRRS (32.13 dB) and Snow100K (34.62 dB) datasets. The proposed method has significant practical applications in autonomous driving and surveillance systems, where accurate snow removal is crucial for maintaining visual clarity.

This work introduces Uni-Removal, an innovative two-stage framework that effectively addresses the critical challenge of domain adaptation in unified image restoration. Contemporary approaches often face significant performance degradation when transitioning from synthetic training environments to complex real-world scenarios due to the substantial domain discrepancy. Our proposed solution establishes a comprehensive pipeline that systematically bridges this gap through dual-phase representation learning. In the first stage, we implement a structured multi-teacher knowledge distillation mechanism that enables a unified student architecture to assimilate and integrate specialized expertise from multiple pre-trained degradation-specific networks. This knowledge transfer is rigorously regularized by our novel Instance-Grained Contrastive Learning (IGCL) objective, which explicitly enforces representation consistency across both feature hierarchies and image spaces. The second stage introduces a groundbreaking output distribution calibration methodology that employs Cluster-Grained Contrastive Learning (CGCL) to adversarially align the restored outputs with authentic real-world image characteristics, effectively embedding the student model within the natural image manifold without requiring paired supervision. Comprehensive experimental validation demonstrates Uni-Removal’s superior performance across multiple real-world degradation tasks including dehazing, deraining, and deblurring, where it consistently surpasses existing state-of-the-art methods. The framework’s exceptional generalization capability is further evidenced by its competitive denoising performance on the SIDD benchmark and, more significantly, by delivering a substantial 4.36 mAP improvement in downstream object detection tasks, unequivocally establishing its practical utility as a robust pre-processing component for advanced computer vision systems.

Transformer-based encoder-decoder models have achieved remarkable success in image-to-image transfer tasks, particularly in image restoration. However, their high computational complexity-manifested in elevated FLOPs and parameter counts-limits their application in real-world scenarios. Existing knowledge distillation methods in image restoration typically employ lightweight student models that directly mimic the intermediate features and reconstruction results of the teacher, overlooking the implicit attention relationships between them. To address this, we propose a Soft Knowledge Distillation (SKD) strategy that incorporates a Multi-dimensional Cross-net Attention (MCA) mechanism for compressing image restoration models. This mechanism facilitates interaction between the student and teacher across both channel and spatial dimensions, enabling the student to implicitly learn the attention matrices. Additionally, we employ a Gaussian kernel function to measure the distance between student and teacher features in kernel space, ensuring stable and efficient feature learning. To further enhance the quality of reconstructed images, we replace the commonly used L1 or KL divergence loss with a contrastive learning loss at the image level. Experiments on three tasks-image deraining, deblurring, and denoising-demonstrate that our SKD strategy significantly reduces computational complexity while maintaining strong image restoration capabilities.