Multi-Class Parrot Image Classification Including Subspecies with Similar Appearance

SIMPLE SUMMARY: Owing to climate change and human overdevelopment, the number of endangered species has been increasing. To face this challenge, the CITES treaty has been adopted by many countries worldwide to prevent the extinction of endangered plants and animals. Additionally, since customs clearance inspections for goods at airports and ports take a long time, and due to the difficulty of distinguishing such species by nonexperts, smugglers have been exploiting this vulnerability to illegally import or export endangered parrot species. If these cases continue to increase, the extinction of species with fewer populations can be accelerated by illegal trade. To tackle this problem, in this study, we constructed an object detection model using convolutional neural networks (CNNs) to classify 11 endangered species of parrots. Utilizing artificial intelligence techniques, the procedures for inspection of goods can be simplified and the customs clearance inspection systems at airports and ports can be enhanced, thus protecting endangered species. ABSTRACT: Owing to climate change and human indiscriminate development, the population of endangered species has been decreasing. To protect endangered species, many countries worldwide have adopted the CITES treaty to prevent the extinction of endangered plants and animals. Moreover, research has been conducted using diverse approaches, particularly deep learning-based animal and plant image recognition methods. In this paper, we propose an automated image classification method for 11 endangered parrot species included in CITES. The 11 species include subspecies that are very similar in appearance. Data images were collected from the Internet and built in cooperation with Seoul Grand Park Zoo to build an indigenous database. The dataset for deep learning training consisted of 70% training set, 15% validation set, and 15% test set. In addition, a data augmentation technique was applied to reduce the data collection limit and prevent overfitting. The performance of various backbone CNN architectures (i.e., VGGNet, ResNet, and DenseNet) were compared using the SSD model. The experiment derived the test set image performance for the training model, and the results show that the DenseNet18 had the best performance with an mAP of approximately 96.6% and an inference time of 0.38 s.