Using Partial Least Squares in Butterfly Species Identification

Butterflies are important insects in nature, and along with moths constitute the Lepidoptera order. At the global level, the number of existing butterfly species is approximately 16,000. Therefore, the identification of their species in images by humans consists in a laborious task. In this paper, we propose a novel approach to recognize butterfly species in images by combining handcrafted descriptors and the Partial Last Squares (PLS) algorithm. A set of PLS models are trained using an one-againstall protocol. The test phase consists in presenting images to all classifiers and the one which provides the highest response value contains in the positive set the predicted class. The performance of the proposed approach is evaluated on the Leeds Butterfly dataset. Experiments were conducted using HOG and LBP descriptors, separately and combined. The approach using HOG singly reported an accuracy rate of 68.72%, while using only LBP resulted in an accuracy rate of 77.33%. Combining both descriptors this value changes to 76.27%. The proposed approach achieves the best results in all three versions when compared to state-of-the-art approaches. Experiments have shown that describing images with LBP provides the highest accuracy values since it extracts texture information, what is an important characteristic to distinguish butterflies. However, information of color and shape, added by HOG, appears to make different species confused.