Accelerating fish detection and recognition by sharing CNNs with objectness learning

Daily increasing underwater visual data makes automatic object detection and recognition a great demand and challenging task. In this paper, we adopt a region proposal network to accelerate underwater object detection and recognition from Faster R-CNN. This process implement detection acceleration by using convolutional networks to generate high-quality object candidates, and by sharing these networks with original detection networks. Applied to a real-world fish dataset, which includes 24,277 ImageCLEF fish images belonging to 12 classes, this automatic detection and recognition system has a nearly real-time frame rate of 9.8 ftps, while yielding 15.1% higher Mean Average Precision (mAP) than the Deformable Parts Model (DPM) baseline.