Relation-IOU: A Novel Bounding Box Regression Loss for Early Apple Disease Detection

Accurate detection of apple leaf diseases at an early stage is essential to help prevent disease spreading and further promote high-quality development of the apple industry. The early apple leaf disease presents with small spots, which lead to difficulties in localizing the disease spots. Therefore, this paper proposed a novel IOU-based regression loss function RelationIOU Loss to assist the model in dealing with the early apple leaf disease detection task. The experimental results established that the Faster R-CNN model with Relation-IOU Loss achieves outstanding performance, with an accuracy of 64.30%, 86.27%, 79.10%, 89.97%, and 65.80% in AP50 for 5 common apple leaf diseases. This established that the proposed Relation-IOU Loss could achieve competitive performance on the early apple leaf disease detection task and satisfy the requirements of large-scale agricultural development.