A Transfer Learning Approach to Bird Species Recognition using MobileNetV2

The growing interest in biodiversity and conservation has heightened the need for accurate and efficient bird species recognition. This is because bird populations, their distribution, migration patterns, and environmental impact assessments can benefit from the valuable data these recognition methods provide. To address the complex task of manual bird species classification, this research paper presents a transfer learning approach based on MobileNetV2. The proposed solution uses MobileNetV2 to categorize bird species in photographs, achieving 96.79% accuracy on a defined dataset. This work highlights the potential of transfer learning in solving complex recognition tasks and the efficiency of MobileNetV2 in achieving high accuracy despite its lightweight design. The method has real-world applications, including real-time bird identification systems and bird population monitoring, and will become more valuable with advancements in hardware and deep learning. To improve the overall system’s accuracy even further, using a larger dataset, scheduling regular retraining sessions and further fine-tuning the model is suggested. In this study, the model is trained separately on two different datasets. The first dataset contains around 1800 images with 200 types of bird species. The second dataset contains around 180 images with 20 types of bird species, specifically South Indian bird species. The attained accuracy with the first dataset is 90.91% and the attained accuracy with the second dataset is 96.79%.