January 18, 2025 | Transactions on CIT |
Researchers from Chiang Mai University and Prince of Songkla University in Thailand conducted a study to address challenges in identifying and diagnosing leaf diseases and pests affecting durian trees. Durian is an important agricultural export for Southeast Asian countries, but its productivity and crop quality are often reduced due to disease and pest infestations, resulting in significant economic losses.
The study focused on using deep learning techniques to classify visual symptoms of diseases and pests from images of durian leaves. However, the development of a high-performing deep neural network requires a substantial number of accurately labeled training images, which are difficult and costly to obtain. To overcome this limitation, the researchers proposed a method to supplement the limited number of expert-labeled images with noisily labeled images gathered from the Internet.
A sample selection framework was introduced to identify and choose useful noisy images to augment the training set. These images were used in a multi-round learning scheme, where each learning round improved the performance of the model. The results showed that incorporating noisy images provided complementary information, increasing prediction accuracy by 20% in one of the learning rounds. The study demonstrates the potential of combining noisy and ground-truth data to enhance disease and pest identification for improved crop management in durian production.
