Synergistic fusion network of GAN-based deblurring and target detection for high-speed potato seeding monitoring
Abstract
• Proposal of a GAN-based detection-guided deblurring framework for high-speed seeding monitoring. • Development of TSF-DeblurGAN with multi-scale feature fusion for enhanced restoration under motion blur. • Introduction of LGCSA-Detector integrating local–global attention for challenging seed detection. • Achievement of robust real-time performance on edge devices under agricultural conditions. Real-time, accurate monitoring of potato miss- and multi-seeding is critical for precision seeding, but challenged by motion blur induced by high-speed seed metering. Severe non-uniform, multi-scale blur causes a substantial loss of semantically discriminative features, rendering conventional “deblur-then-detect” pipelines ineffective. To overcome this, we propose a novel end-to-end GAN (Generative Adversarial Network)-based deblurring–target detection (GDTD) framework, underpinned by a synergistic training mechanism that jointly optimizes image restoration and object detection. Evaluated under high-speed potato seeding conditions with dynamic non-uniform blur, The GDTD achieved an [email protected]:0.95 (mean Average Precision) of 0.6843 and a Precision of 0.9115, outperforming two-stage fusion frameworks built on state-of-the-art deblurring (TSF-DeblurGAN) and detection (YOLO-based LGCSA-Detector) models by up to 19.7% and 5.73%, respectively. Furthermore, it sustained real-time inference at a processing speed of 27.3 FPS with low power consumption, making it suitable for embedded agricultural devices. By shifting the optimization paradigm from pixel-level fidelity to task-aware semantic restoration, this study provides important theoretical and practical support for precision seeding and other dynamic scenarios in smart agriculture, demonstrating that task-driven semantic restoration, rather than pixel-perfect deblurring, is the key to robust detection under extreme motion blur.