EfficientNet: A Low-bandwidth IoT Image Sensor Framework for Cassava Leaf Disease Classification
Chih-Cheng Chen, Ju Yan Ba, Tie Jun Li, Christopher Chun Ki Chan, Kun Ching Wang, Zhen Liu
Abstract
Following the cassava leaf disease classification process, we successfully design a novel convolutional neural network (CNN) framework called EfficientNet using low-bandwidth image sensors and a combination of image enhancement and image classification methods. For this study, we have employed low-bandwidth, small-scale IoT image sensors in a farm to capture images of cassava leaves at periodic intervals. We employ data enhancement techniques such as test-time augmentation (TTA) and cutmix, cutout, and k-fold to accurately classify and evaluate the pathology of cassava plants. We carry out multiple simulated experiments to classify and evaluate diseases found in five cassava leaf datasets. Our framework is capable of producing relatively accurate classification results despite small differences between the test set images, and we achieved a classification accuracy for our final test set of 89%, comparable with that in similar studies. Experimental results obtained using a sensor show that EfficientNet significantly outperforms a state-of-the-art cassava leaf disease classification model.