AiOENet: All-in-One Low-Visibility Enhancement to Improve Visual Perception for Intelligent Marine Vehicles Under Severe Weather Conditions.

Benefiting from the higher performance-cost ratio and installation convenience, the visible-light imaging camera has become one of the most widely-used onboard sensor for safe vehicle navigation. However, the captured images inevitably suffer from the color distortion, contrast reduction, or loss of fine details under severe weather conditions (such as haze, low-lightness, rain, and snow). The quality-degraded visual information will lead to the limited perceptual accuracy and range, resulting in the increased navigation risk for intelligent marine vehicles. To suppress the influences of severe imaging conditions on navigation safety, this work proposes an all-in-one low-visibility enhancement network (termed AiOENet) to improve the visual perception for marine surface vehicles under different weather scenarios. In particular, our AiOENet mainly consists of a VGG16-driven scene discriminator, an encoder, parameter-shared Transformer blocks, and a decoder. The scene discriminator is exploited to classify four types of low-visibility images, collected under the hazy, low-light, rainy, or snowy weather conditions. According to the classification results, the low-visibility images are fed into the corresponding encoders for coarse feature extraction. The multiple Transformer blocks are then employed to separate and extract the smaller-scale features. The latent normal-visibility images are finally generated through the corresponding decoders. Therefore, our AiOENet has the capacity of flexibly and adaptively restoring the diverse low-visibility images using a uniform encoder-decoder network architecture under different imaging conditions. Compared with the state-of-the-art imaging methods, AiOENet achieves the comparable or even superior enhancement results in terms of both quantitative and qualitative evaluations. In addition, the AiOENet can contribute to the more accurate and stable object detection with the improved visual perception in maritime low-visibility scenes. The source code is available at https://github.com/LouisYuxuLu/AiOENet .