Marine Bird Detection Based on Deep Learning using High-Resolution Aerial Images

The overarching goal of this paper is to find an automatic bird detection and counting method on aerial images of the ocean. Most of the existing works in the literature are based on heuristic handcrafted feature design, which in most cases affect the effectiveness (the accuracy of classification) and the efficiency (spending much time). In this paper, we propose a method built on a systematic feature learning based classification adopting a new deep Convolutional Neural Network (CNN) architecture. Through this architecture, the feature learning is automated from a multi-dimensional raw input images, by a training step leveraging the JONATHAN dataset via supervised learning. Performances evaluation show that the CNN based architecture classifier achieves an accuracy of 95% on the JONATHAN test data and the overall detection method achieves a classification rate (true positives: birds) of 98% on real images.