Leveraging Multi-Round Learning and Noisy Labeled Images from Online Sources for Durian Leaf Disease and Pest Classification

Abstract

Durian has recently become a major agricultural export commodity for Southeast Asian countries. However, this plant is vulnerable to various diseases and pests, which are usually considered the main cause of poor yields and low-quality crops; leading to a huge economic loss. This work focuses on leaf diseases and pests, as their symptoms can be easily detected visually, but it is still challenging to correctly diagnose the problems. To address this difficulty, this work makes use of deep learning techniques to classify a given photo to a corresponding class of disease or pest. However, building a high-performance deep neural network model requires a substantial amount of ground-truth photos of diseases and pests on durian leaves, which are difficult and expensive to acquire. To overcome this challenge, we propose enriching the limited number of expert-labeled images with abundantly available, noisily labeled images collected from the Internet. A sample selection framework is introduced to choose noisy images for augmenting a current training set, which will be used to build a new classifier in the next learning round. We found that such a multi-round learning scheme, in which noisy photos are intuitively selected, provides complementary information to the limited ground truth, thereby enhancing the prediction accuracy on unseen examples of a classifier being built by 20% at a particular learning round.