EG-ISNet: An Edge-Guided Image Stitching Network for Drone-Captured Ship Hull Images

Ships, which operate in marine environments, are susceptible to various damages, resulting from seawater corrosion, wind and wave impacts, and biofouling. Traditional manual hull inspection operations are often inefficient and hazardous. Recent advancements in drone techniques have facilitated hull image acquisition, which produces a large number of images of the hull surface for inspection. However, a significant challenge in ship hull image stitching is attributed to the performance degradation under low-texture, low-overlapping and large-parallax conditions. To address this challenge, we propose an Edge-Guided Image Stitching Network (EG-ISNet). This network combines an edge detection sub-network with an image registration sub-network, forming two parallel branches. The EG-ISNet can be trained using a multi-task learning scheme. As a result, the ability of the registration sub-network to capture edge characteristics is enhanced, which increases the accuracy of the registration operation. The registered images are subsequently processed through a weighted average fusion operation. The result is the final stitched hull image. Our network is tested using a set of drone-captured hull images. We evaluated our network on a set of drone-captured hull images, and the results demonstrate its superiority over six existing methods in scenarios with low texture and significant parallax.