Detection and localization for lake floating objects based on CA-faster R-CNN

As a general trend, unmanned ships have been gradually replacing humans and served as the cleaner of lakes. To work properly, those unmanned ships need to detect and localize lake floating objects that need to be collected. Compared to conventional image-based objects, lake floating objects are too small to detect. Meanwhile, because most conventional algorithms depend on bounding-boxes to detect the object, their results - it is hard to detect the accurate location of floating objects. To this end, this paper proposes a detection and localization algorithm based on CA-Faster R-CNN (Class Activation-Faster Regions with Convolutional Neural Network). Specifically, for an image with objects on it, the proposed algorithm detects and classifies objects with Faster R-CNN and localize objects with CA network. The experimental results show that, compared with the Faster R-CNN algorithm, this algorithm can reduce the positioning error without affecting the recognition accuracy, thereby can be used for the detection and localization of floating objects on the water surface. Compared with Faster R-CNN algorithm, the positioning accuracy of CA-Faster R-CNN algorithm is improved by 6.29 pixels. Also, the proposed algorithm remains a great potential for other objects that shared similar challenges with lake floating objects.