Key Area Acquisition Training for Practical Image-based Plant Disease Diagnosis

Automatic diagnosis of plant diseases using images is a fine-grained task, and disease symptoms are often ambiguous and highly variable. Pre-extraction of the region of interest (ROI) exhibiting disease symptoms (such as one or more leaves) is known to have a certain effect on improving accuracy. However, the ROI extraction at runtime is time-consuming, resulting in issues of system usability. This paper proposes a new training method called key area acquisition training (KAAT). KAAT reduces the variation in prediction results between images before and after the extraction of the ROI. By directing the model’s attention to the ROI through learning, KAAT contributes to improved diagnostic performance without sacrificing execution time during diagnosis. In the evaluation, we conducted nine class diagnosis task (eight diseases and healthy) using 77K and 9K images of cucumber leaves (collected from different fields) for training and testing, respectively. The proposed KAAT improved diagnostic accuracy by 3.8% in macro-F1 and 2.0% in micro accuracy without increasing execution time.