RL-RiceBox: A Reinforcement Learning Framework for Bounding Box Refinement in RiceLeaf Disease Localization

Abstract

Precise localization of diseased regions on rice leaves is critical for reliable diagnosis, yet bounding-box annotations and detector outputs are often coarse and inconsistent. We propose RL-RiceBox, a deep reinforcement learning framework that automatically refines an initial (inaccurate) bounding box through a short sequence of discrete geometric actions (translation, scaling, and termination). The refinement process is formulated as a Markov Decision Process, and a Deep Q-Network learns an action policy from a fixed cropped patch together with the current box coordinates, using an IoU-driven reward to encourage fast convergence to tighter lesion localization. To emulate realistic initialization noise, training and evaluation start from synthetically shifted boxes generated from the ground truth. Experiments on a rice leaf disease dataset show that RL-RiceBox consistently improves localization quality, increasing mean IoU from 0.53 to 0.89 and achieving 90.4% and 88.6% success rates at IoU thresholds of 0.50 and 0.75, respectively, with an average of 7.1 refinement steps. The proposed approach provides an effective post-processing module for improving box quality and strengthening agricultural vision pipelines.