Vision-based ship segmentation in bridge areas using attention-guided multiscale region searching approach

Ship identification is an important part of load monitoring in bridge health monitoring systems for cross-river or cross-sea bridges. Presently, increasing concern is being paid by worldwide bridge authorities to ship-bridge collision risk reductions. This paper proposes a vision-based framework for ship segmentation using a presented attention-guided multiscale region searching approach using ship monitoring images (especially high-resolution images) in bridge areas. The ship searching approach consists of two steps: (1) attention-guided multiscale sampling technique, which employs low-rank and sparse decomposition method for candidate region filtering to improve the efficiency; and (2) ship region detection technique, which is aiming to eliminate the ineffective candidate regions and completed by a single-shot convolutional neural network for precise region detection. Eventually, accurate ship contours are extracted by the proposed corner-point-based convex hull finding (CP-CHF). Discussions about the efficiency of the sampling technique and CP-CHF are also presented. The application shows the accuracy and consistency of the proposed ship segmentation framework.