Degradation-adaptive neural network for jointly single image dehazing and desnowing

In harsh winter scenes, the captured images often suffer from haze and snow degradations simultaneously, which significantly affect the performance of high-level computer vision tasks. Most existing restoration methods are either specialized to address only one type of weather-related degradation, or have a heavy number of parameters.

To address these problems, a research team led by Erkang Chen and Yun Liu conducted a study, now published in Frontiers of Computer Science.

The team proposed a lightweight image restoration network called Degradation-Adaptive Neural Network (DAN-Net) to achieve jointly single image dehazing and desnowing. DAN-Net is verified and tested on five large-scale datasets.

Compared to existing research results, the proposed method guarantees superior performance with lower computational complexity for the dehazing and desnowing tasks.

In the research, they carefully design the overall architecture of DAN-Net, which consists of two task-specific expert networks and an adaptive gated neural network. The task-specific expert network is constructed using three high-performance convolution-based components (i.e., MSTB, DPAM and CLAGM) and three-level layers.

Unlike previous learning-based methods that simply stack the components repeatedly, the proposed expert network embraces lightweight advantage and achieves the excellent parameter-performance trade-off through careful design in each layer and the usage of few components.

The adaptive gated neural module is developed as an effective degradation-adaptive guider to control the contributions of two pre-trained expert networks.

The experiments are performed on five large-scale datasets including RESIDE, Haze4k, CSD, SRRS and Snow 100K. Experimental results demonstrate the superiority of the proposed DAN-Net with lower computational complexity on both synthetic and real-world degraded images under winter scenes.

In future work, the researchers will focus on exploring the design of task-specific expert networks and the adaptive gated neural module to improve the performance and efficiency of DAN-Net.

In harsh winter scenes, the captured images often suffer from haze and snow degradations simultaneously, which significantly affect the performance of high-level computer vision tasks. Most existing restoration methods are either specialized to address only one type of weather-related degradation, or have a heavy number of parameters.

To address these problems, a research team led by Erkang Chen and Yun Liu conducted a study, now published in Frontiers of Computer Science.

The team proposed a lightweight image restoration network called Degradation-Adaptive Neural Network (DAN-Net) to achieve jointly single image dehazing and desnowing. DAN-Net is verified and tested on five large-scale datasets.

Compared to existing research results, the proposed method guarantees superior performance with lower computational complexity for the dehazing and desnowing tasks.

In the research, they carefully design the overall architecture of DAN-Net, which consists of two task-specific expert networks and an adaptive gated neural network. The task-specific expert network is constructed using three high-performance convolution-based components (i.e., MSTB, DPAM and CLAGM) and three-level layers.

Unlike previous learning-based methods that simply stack the components repeatedly, the proposed expert network embraces lightweight advantage and achieves the excellent parameter-performance trade-off through careful design in each layer and the usage of few components.

The adaptive gated neural module is developed as an effective degradation-adaptive guider to control the contributions of two pre-trained expert networks.

The experiments are performed on five large-scale datasets including RESIDE, Haze4k, CSD, SRRS and Snow 100K. Experimental results demonstrate the superiority of the proposed DAN-Net with lower computational complexity on both synthetic and real-world degraded images under winter scenes.

In future work, the researchers will focus on exploring the design of task-specific expert networks and the adaptive gated neural module to improve the performance and efficiency of DAN-Net.