Multi-Condition Training on Deep Convolutional Neural Networks for Robust Plant Diseases Detection

Currently, deep learning has become prominent technology for many computer vision tasks. However, while deep learning achieves satisfactory performance when evaluated with data from identical distributions, it still suffers when data are distorted. In many cases, it would be difficult and costly to collect data on all possible conditions. To deal with it, we propose Multi-condition training (MCT) to train more robust deep convolutional neural networks. MCT works by corrupting the existing “ideal” data with various possible environmental conditions. We evaluated the method on tea diseases dataset. Our results confirms that MCT produces more robust deep learning systems and is shown effective to deal with distorted images.