Lightweight GAN for Restoring Blurred Images to Enhance Citrus Detection

Abstract

Image blur is a major factor that degrades object detection in agricultural applications, particularly in orchards where crop occlusion, leaf movement, and camera shake frequently reduce image quality. This study proposed a lightweight generative adversarial network, AGG-DeblurGAN, to address non-uniform motion blur in citrus tree images. The model integrates the GhostNet backbone, attention-enhanced Ghost modules, and a Gated Half Instance Normalization Module. A blur detection mechanism enabled dynamic routing, reducing computation on sharp images. Experiments on a citrus dataset showed that AGG-DeblurGAN maintained restoration quality while improving efficiency. For object detection, restored citrus images achieved an 86.4% improvement in mAP@0.5:0.95, a 76.9% gain in recall, and a 40.1% increase in F1 score compared to blurred images, while the false negative rate dropped by 63.9%. These results indicate that AGG-DeblurGAN can serve as a reference for improving image preprocessing and detection performance in agricultural vision systems.