Structure-Driven Feature Modeling for Robust Fish Detection in Complex Underwater Environments

Abstract

Simple Summary Accurate fish monitoring is important for aquaculture management and animal observation, but detecting fish in underwater images is difficult when they are small, blurred, partly hidden, poorly illuminated, or visually similar to the background. This study developed an image-based detection method that places greater emphasis on fish boundaries, shapes, and body structure while retaining useful appearance information. The method was evaluated using a newly constructed dataset of Medaka, also known as Japanese rice fish, collected under challenging rearing conditions. The proposed method improved detection accuracy compared with the original detection model and produced consistent improvements when applied to several related detection models. The method was also evaluated on a public underwater fish dataset containing four different fish species, where it again improved detection performance compared with the original model. These results indicate that emphasizing structural information can help computers recognize fish more reliably when their visual appearance is unclear or partly obstructed. The proposed method may support more accurate and efficient fish monitoring, reduce the need for continuous manual observation, and provide useful technical assistance for aquaculture management and studies of fish behavior.