Spatial Feature Selection for Underwater Object Detection

Due to the high error detection rate in the underwater identification task, we introduce multi-scale feature selection and center and boundary feature selection for a fully convolutional one-stage object detector(FCOS)to make the high-quality underwater object detection. In the adaptive weighted fusion feature pyramid, multi-scale spatial feature selection can be realized by setting learnable weights to integrate all feature levels. In addition, in order to solve the feature coupling entanglement between the classification and regression task, and separate the shared features among different tasks, a feature decoupling detection head based on spatial feature was designed to realize the feature selection of center and boundary regions. Experiments are carried out on the underwater datasets URPC2018 and UWD2021, comparing with other object detection methods. Extensive experiments results demonstrate that the proposed method shows extremely excellent performance, achieving 82. 7% mAP and 83. 3% mAP on URPC2018 and UWD2021.