MDD-ShipNet: Math-Data Integrated Defogging for Fog-Occlusion Ship Detection

For maritime autonomous surface ships, challenges exist in visual detection of ships in sea foggy scenarios, thereby severely degrading visual detection autonomy. In this paper, math-data integrated defogging (MDD) mechanism is created within a ship detection network, termed MDD-ShipNet. Main contributions are as follows: 1) The MDD enhancement module (MDD-EM) is implemented by devising 5 filters, i.e., defog, exposure, tone, contrast and sharpen, as well as a CNN-based parameter learner, such that defogging enhancement can progressively be conducted in a transparent manner; 2) The detector is innovated by employing polarized self-attention (PSA) and weighted bidirectional feature pyramid network (WBiFPN), so as to preserve long-range dependencies and high-resolution channel-spatial features, simultaneously, thereby sufficiently fusing shallow and semantics information associated with contributions to detection; and 3) The entire MDD-ShipNet framework is ultimately established in a weakly supervised manner by integrating MDD-EM and PSA-WBiFPN-based detector, and is fertilized by hybrid dataset that is diversely contributed by real-world and synthesized sea-foggy images. Comprehensive experiments and comparisons eventually validate that the MDD-ShipNet framework outperforms typical approaches deriving from the detection after image enhancement, multi-task learning and domain adaption in terms of mAP@.5, mAP@[.5:.95] and FPS.