Performance Analysis of Deep Learning and Machine Learning for Bird Species Recognition

The study of bird species classification plays a major role in the conservation of biodiversity. In the present work, we have presented a performance analysis of various deep learning and machine learning techniques, along with their ensembles, for bird species classification based on images. This helps identify the most reliable methods for tracking bird populations. The study utilizes Kaggle’s repository of bird images from 20 species, consisting of 3,208 training images, 100 test images, and 100 validation images. Additionally, the Caltech 200 2011 dataset, which contains 11,788 images grouped into 200 classes, is also used. Deep feature extraction methods, such as MobileNet with transfer learning, as well as traditional machine learning feature extractors for shape, color, and texture features, and their combinations, have been employed in this work. Various deep architectures are used for the automatic extraction of features and classification. Different classifiers, including Random Forest, Bagging, SVC, Ridge, Decision Tree, K-Nearest Neighbor, and their ensembles, were utilized for classification. The SVC classifiers provided the highest performance when MobileNet with transfer learning was used for feature extraction, outperforming traditional machine learning techniques. In conclusion, deep learning architectures with fully automated feature extraction and classification have outperformed all existing methods, demonstrating that deep learning surpasses traditional machine learning methods in bird species classification.