Automatic segmentation and three-dimensional reconstruction system for monocular images of megabenthos based on manned submersible

Abstract Accurate analysis of the three-dimensional (3D) morphological characteristics of megabenthos is crucial for the study and protection of deep-sea ecosystems. However, existing 3D reconstruction methods based on laser and depth cameras are expensive and unwieldy. Furthermore, the deep-sea imaging environment is complex, with significant background interference that is difficult to distinguish from the megabenthos, affecting the accuracy and speed of reconstruction. To address these issues, an automatic segmentation and 3D reconstruction system for monocular images of megabenthos based on manned submersible was proposed. The monocular vision system of the manned submersible was used as the image acquisition platform by the system. First, a biological image segmentation model based on edge prior optimized UNet was designed to accurately segment the biological region from the two-dimensional monocular image, avoiding the impact of the complex background on the reconstruction. Second, a biological 3D reconstruction model based on fractal enhanced Gaussian splatting was designed to achieve 3D reconstruction of the megabenthos inexpensively and effectively. Finally, compared with advanced segmentation and reconstruction models, the system’s performance is superior. The morphological characteristics of a megabenthos in a seamount in the western Pacific Ocean were analyzed using this system. The results demonstrate that the system achieves accurate segmentation and high-precision reconstruction within a unified workflow, enabling automated extraction and quantitative analysis of 3D morphological features of deep-sea organisms.