FlexIoU: Modularity and Flexibility IoU for Bounding Box Regression in Agricultural Disease Detection

Abstract

Accurate bounding-box regression plays a crucial role in object detection, particularly in agricultural disease detection, where target objects are often small, irregular, and weakly bounded. Although Intersection over Union (IoU)-based losses such as CIoU and EIoU have improved localization performance, their fixed geometric penalty formulations limit adaptability to domain-specific object characteristics. In this work, we propose FlexIoU, a flexible and lightweight IoU-based regression loss that introduces tunable geometric penalties to enhance localization robustness. FlexIoU provides explicit control over center-distance, width, and height penalties, enabling more adaptive regression behavior for irregular lesion shapes commonly observed in agricultural imagery. The proposed loss integrates seamlessly into the YOLO training pipeline as a drop-in replacement for conventional IoU-based losses, without modifying classification or distribution focal loss components, thereby preserving inference efficiency. We evaluate FlexIoU within a YOLOv11 detection framework on a tomato leaf disease dataset under identical experimental settings and across three random seeds to ensure statistical reliability. Experimental results demonstrate that FlexIoU consistently outperforms the default YOLOv11 regression objective as well as CIoU and EIoU in terms of mAP@0.5:0.95 and overall detection balance, while maintaining stable inference speed comparable to baseline losses. These findings indicate that FlexIoU provides an effective and practical solution for improving bounding-box regression in agricultural disease detection. While the results demonstrate promising robustness across two tomato disease datasets, further validation on additional agricultural and general object detection datasets is required.