Benefiting from video surveillance systems that provide real-time traffic conditions, automatic vessel detection has become an indispensable part of the maritime surveillance system. However, high-level vision tasks generally rely on high-quality images. Affected by the imaging environment, maritime images taken under poor lighting conditions easily suffer from heavy noise and colour distortion. Such degraded images may interfere with the analysis of maritime video by regulatory agencies, such as vessel detection, recognition and tracking. To improve the accuracy and robustness of detection accuracy, we propose a lightweight generative adversarial network (LGAN) to enhance maritime images under low-light conditions. The LGAN uses an attention mechanism to locally enhance low-light images and prevent overexposure. Both mixed loss functions and local discriminator are then adopted to reduce loss of detail and improve image quality. Meanwhile, to satisfy the demand for real-time enhancement of low-light maritime images, model compression strategy is exploited to enhance images efficiently while reducing the network parameters. Experiments on synthetic and realistic images indicate that the proposed LGAN can effectively enhance low-light images with better preservation of detail and visual quality than other competing methods.

Maritime video surveillance has become an essential part of the vessel traffic services system, intended to guarantee vessel traffic safety and security in maritime applications. To make maritime surveillance more feasible and practicable, many intelligent vision-empowered technologies have been developed to automatically detect moving vessels from maritime visual sensing data (i.e., maritime surveillance videos). However, when visual data is collected in a low-visibility environment, the essential optical information is often hidden in the dark, potentially resulting in decreased accuracy of vessel detection. To guarantee reliable vessel detection under low-visibility conditions, the paper proposes a low-visibility enhancement network (termed LVENet) based on Retinex theory to enhance imaging quality in maritime video surveillance. LVENet is a lightweight deep neural network incorporating a depthwise separable convolution. The synthetically-degraded image generation and hybrid loss function are further presented to enhance the robustness and generalisation capacities of LVENet. Both full-reference and no-reference evaluation experiments demonstrate that LVENet could yield comparable or even better visual qualities than other state-of-the-art methods. In addition, it takes LVENet just 0⋅0045 s to restore degraded images with size 1920 × 1080 pixels on an NVIDIA 2080Ti GPU, which can adequately meet real-time requirements. Using LVENet, vessel detection performance can be greatly improved with enhanced visibility under low-light imaging conditions.