Detection of Harbour Porpoise with Low-Level Feature Extraction and Deep Learning Based Classification

This paper proposes a method for detection of harbour porpoise from very high resolution images. The approach involves the detection of salient objects of very small sizes with the feature accelerated segment test method, creation of an annotated database and classification of those objects with different convolutional neural networks. Multiple deep models based on VGGNet, ResNet, and GoogLeNet are modified and used in this work while a shallow custom network is developed as well. This method is computationally less expensive because of the network application on smaller candidate patches instead of the entire image. Provided the versatile and challenging environmental conditions during the period of the video recording, the networks show a good and robust performance.