Effect of Spatial Dropout on Mosquito Classification using VGGNet
Katanyu Tharawatcharasart, Wanchalerm Pora · 2022 19th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON) · 2022
Monitoring systems with automatic classification have been developed for years but never reached the human expert level. Convolutional neural networks (CNN) achieve state-of-the-art results in image classification tasks and can be applicable for real-time image classification. However, CNN usually may be overfitting during training due to relatively small datasets. In this work, we investigate the effects of transfer learning, spatial dropout, and image augmentation techniques in the VGGNet architectures. Our results show that VGG-16 achieves more than 96.9% of accuracy when proper techniques are applied to the networks. Moreover, combining the predictions from VGG-16 and VGG-19 increases accuracy to 97.2% on the test dataset.