Underwater target detection method based on spatial feature selection

Summary: Aiming at the high false detection rate of traditional target detection methods in underwater recognition tasks, based on a one-stage full convolution detector(FCOS)Introduce multi-scale feature selection and center boundary feature selection to achieve high-precision underwaterTarget Detection.The adaptive weighted fusion feature pyramid in the model achieves multi-scale spatial feature selection by setting learnable weights to weight and fuse all feature levels.Furthermore, in order to deal with the feature coupling problem between classification and regression tasks in detection, and to separate the shared features between different tasks, a detection head based on spatial feature decoupling is designed to realize feature selection in the center and boundary regions.In the experiment, for underwater datasetsURPC2018andUWD2021Perform performance tests and compare with advanced object detection methods.Extensive experimental results show that the FCOSmodel based on spatial feature selection is used for underwater detection tasks.showed excellent performance in URPC2018andUWD2021