Classifying Bird Feeder Photos

Orange Fronted Parakeets are New Zealand native birds which are critically endangered. The Department of Conservation has a dedicated team to keep these birds protected. Their work includes manually classifying bird-feeder photos to investigate which parakeets visit the feeders. This paper proposes a method to classify bird-feeder images into parakeets, other birds and empty feeders.Two methods were proposed, both of which used a Convolutional Neural Network (CNN) to classify images. These were a single-pipe multiclass classification and a two-pipe classification. The single pipe classification determined whether the image contained a parakeet, other bird or was empty. The two-pipe method classified the image into bird and empty and then classified the image into other bird and parakeet. The proposed model was pretrained on the ImageNet database using ResNet18. This model was then fine-tuned on the bird feeder data base. Each image underwent normalization, a random resized crop and a random horizontal flip. The testing true positive accuracy of the single-pipe model was 96.53% accurate. While the two-pipe model achieved an accuracy of 97.99%. Both these models outperformed the previous model which had an accuracy of 89%.