Highly Robust Dolphin Detection Algorithm in Occluded Cases

Marine mammal detection is helpful for ecological conservation. In this paper, we proposed a novel automatic method in real-time to detect and recognize the dolphin that is underwater and just reveals some characteristics of the body. The proposed method is based on the convolutional neural network with an extra masking layer to approximate various filter without knowing the normal neural network, which explore a discriminative criterion to enhance the image segmentation performance. We evaluate the proposed system on our on-site dolphin shooting datasets. The proposed approach can achieve higher detection and recognition rates than existing classification models and outperform faster R-CNN baselines in the object detection.