基于改进Faster R-CNN的水面漂浮物识别与定位

Detection and Localization of Floating Objects via Improved Faster R-CNN

为了解决智能无人船水面漂浮物识别和定位精度不高的问题,提出了一种基于Faster R-CNN(Faster Regions with Convolutional Neural Network)的改进识别与定位算法(CA-Faster R-CNN).该方法采用Faster R-CNN算法对水面漂浮物进行初次识别和定位,对输出的识别结果与定位框采用类别激活网络(Class Activation,CA)去除边界框,运用像素点来标注目标位置.实验结果表明,该算法具有较高的识别与定位精度,可用于水面漂浮物识别和定位.此外,该算法对于其他与水面漂浮物具有相似特征的小目标物体定位有一定的借鉴作用.

A detection and localization algorithm was proposed to overcome the issue that unmanned ships has low accuracy in detecting and locating floating objects.The proposed algorithm utilized Faster R-CNN(Faster Regions with Convolutional Neural Network)as the building block to conduct the initial detection and localization,output the temporal result with location boxes.Then the Class Activation network(CA)was used to remove the location boxes and mark the object location with pixels.The case study has verified that the algorithm has satisfactory accuracy in detection and localization of floating objects.Besides,the algorithm remains potential for other objects that shared similar features with lake floating objects.