Exploring Fine-Grained Feature Analysis for Bird Species Classification using Layer-wise Relevance Propagation

This research investigates fine-grained feature analysis in bird species classification using the CUB-200-2011 dataset, specifically focusing on Layer-wise Relevance Propagation (LRP). The dataset’s annotated bounding boxes and part annotations offer a unique opportunity to explore the contribution of intricate details in categorizing bird species. We propose tailored LRP models to effectively analyze fine-grained features in bird images. Our study involves an in-depth feature attribution analysis using LRP to elucidate the neural network models’ decision-making process. We aim to identify the features or parts of bird images that play a significant role in classification, thereby providing insights into the discriminative aspects crucial for accurate categorization. Through LRP techniques, we localize and high-light discriminative features within bird images, such as feather patterns, shapes, or colors, contributing to the advancement of fine-grained classification in ornithology.