A Study on Coral Species Recognition Using Repetitive Structures and Deep Learning

Object detection is used to recognize and locate objects within images or videos. Coral reefs exhibit the highest levels of biodiversity. Typically, corals consist of repetitive structures. In this study, we detected the repetitive structures of corals and then applied the watershed algorithm to segment the entire coral colony using them. This approach enabled comprehensive coral detection and solved related issues. The results indicated that the detection performance for individual coral structures was higher than for the entire colony. Individual structures were used to train the algorithm by selecting individual repetitive coral structures in the same number of frames. Moreover, the detected bounding boxes were primarily applied to repetitive structures exhibiting minor appearance variations to significantly improve the accuracy and recall rates. In selecting individual repetitive coral structures, smaller numerical values were used than in selecting the entire colony. In terms of the resulting images, using individual repetitive coral structures allowed for avoiding non-target objects and produced bounding boxes. The proposed method effectively identified the target corals and accurately delineated the occluded regions when encountering occlusion issues. This approach has the potential for practical application to other coral species or organisms with repetitive structures.