基于空间特征选择的水下目标检测方法

Spatial Feature Selection for Underwater Object Detection

针对传统目标检测方法在水下识别任务中误检率较高的问题,基于一阶段全卷积检测器(FCOS)引入多尺度特征选择及中心边界特征选择,实现高精度水下目标检测。模型中的自适应加权融合特征金字塔通过设置可学习权重加权融合所有的特征层级,实现多尺度空间特征选择。此外,为了处理检测中分类和回归任务之间的特征耦合问题,并分离不同任务之间的共享特征,设计了基于空间特征解耦的检测头,实现了中心和边界区域的特征选择。实验中,针对水下数据集URPC2018和UWD2021进行性能测试,并与先进的目标检测方法进行对比。大量的实验结果表明,基于空间特征选择的FCOS模型在水下检测任务中展现出优异的性能,在URPC2018和UWD2021上的类平均精度(mean Average Precision,mAP)分别为82.7%和83.3%。

Due to the high error detection rate in the underwater identification task,we introduce multiscale 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% m AP and 83.3% m AP on URPC2018 and UWD2021.