Nature Meets Technology: Revolutionizing Butterfly Classification with MobileNetV2
Arpanpreet Kaur · 2024
Belonging to the order Lepidoptera, butterflies are essential ecological indicators and pollinators; nonetheless, their hand identification is difficult because of minute wing pattern differences. This work aims to build an automated butterfly classification system with MobileNetV2, a lightweight convolutional neural network architecture fit for mobile devices. Achieved via depthwise separable convolutions and an inverted residual structure, MobileNetV2's efficiency qualifies it for real-time field conditions categorization. Originating from Kaggle, the input dataset comprises 9,285 training photos and 375 images apiece for validation and testing, therefore covering 75 butterfly species. Techniques of pre-processing and augmentation were used to enhance model training so guaranteeing the diversity and quality of the dataset. The approach consisted in optimizing MobileNetV2 to fit butterfly-specific characteristics and then model evaluation depending on accuracy and other performance criteria. The results show a great classification accuracy of 94.6%, therefore verifying MobileNetV2's ability to precisely identify between species. This success emphasizes how well deep learning might monitor and protect biodiversity. The light weight architecture makes it possible to install on mobile devices, therefore enabling field real-time species identification. Future studies could increase the dataset and investigate different designs to improve the performance and applicability of the model in ecological investigations.