Heterogeneous Twin Dehazing Network for Visibility Enhancement in Maritime Video Surveillance

Visible images captured by maritime video surveillance system under hazy environment often suffer from low contrast and information loss, etc. These negative factors will significantly limit the development and progression of target recognition, detection, tracking, and other vision-based technologies in maritime applications. However, existing dehazing methods are mainly designed for massive inland images, resulting in inaccurate estimation of haze-free versions under the maritime scene. In this work, we propose a heterogeneous twin dehazing network (termed HTDNet) for enhancing visually-degraded maritime images. The HTDNet is mainly composed of two modules, i.e., twin feature extraction module (T-FEM) and feature fusion module (FFM). In particular, T-FEM is employed to collect coarse haze features from two views, and FFM is designed for feature integration and enhancement. To further improve the visual quality, the dataset containing massive maritime images is constructed to train our network. Both synthetic and real-world experiments have illustrated our superior performance compared with several state-of-the-art methods. There is thus a great potential to successfully extend our HTDNet to enable maritime intelligent transportation system.