Region based object detectors for recognizing birds in aerial images

Abstract

This project explores different types of deep neural networks (DNNs) for recognizing birds in aerial images based on real data provided by the Missouri Department of Conservation. The pipeline to identify birds from an image consist of two phases. First, region proposals are created by a DNN, where each region proposal is a sub-area of the image that possibly contains a bird. Second, a DNN is trained as a bird classifier using these region proposals. The bird detection performance is evaluated using the Precision, Recall and F1 scores on a separate test dataset. For the region proposal phase, a Region Proposal Network (RPN) has been implemented and tested, obtaining a Recall above 0.99, which means that the region proposal boxes cover almost all the birds. For the classification phase, a modification of Fast Region-based Convolutional Neural Network (Fast RCNN), a simple Convolutional Neural Network (CNN), and a Capsule Network, have been implemented and tested. For all of them, different hyper-parameters have been explored to increase the final F1 score. These models have been evaluated using two bird dataset variants: easy (with simple backgrounds) and hard (with complex background). Experimental results show that birds can be effectively recognized using the DNNs, especially in the easy dataset. Fast RCNN with a backbone architecture of ResNet50 and in conjunction with other techniques like Feature Pyramids Networks achieved the best results, with a maximum F1 score of 0.902. Simple CNN and Capsule Network achieved a score slightly above 0.8. The techniques used, datasets and results are analyzed to find the main causes of failures in some situations.