An Improved Object Recognition Algorithm via Feature Point: A Case Study of Fish Contour Recognition in Marine Remote Sensing

The ocean archives massive, stable remote sensing datasets, and leveraging these data to achieve intelligent real-time recognition of marine organisms has become a core task in the field of marine remote sensing. However, existing object recognition algorithms primarily focus on determining the location of objects, neglecting the demand for refined object recognition. When these algorithms that only identify location and category applied to marine remote sensing, they fail to meet the requirements of marine fisheries, species conservation, and precise underwater operations. To expand the application scenarios of marine target recognition, we propose a Key Point Refined Network (KPR-Net) for fish contour recognition, a two-stage adaptive target detection algorithm that first determines the positions of target feature points through feature extraction and sufficient fusion processing, then delineates the target contour via simple topological relationships. Our proposed KPR-Net efficiently and accurately predicts the positions of marine target feature points by integrating a self-attention mechanism with multi-feature aggregation. Furthermore, to incorporate biological characteristics and enhance the detection performance of fish targets, we take the natural contour features of fish as constraints, accurately determine the arrangement of these vertices based on the positions and directions of the target feature points, and sequentially connect them according to the determined arrangement to obtain the target contour. Experiments on multiple challenging marine datasets demonstrate the accuracy of our proposed method in multi-category target detection tasks. Particularly in the refined presentation stage of recognition results, the determination of target contours enriches the output content of the target detection algorithm, providing more detailed and comprehensive recognition information.