ReSynDet: A Generate–Assess–Train Framework for Reliable Use of Synthetic Images in Data-Scarce Brassica chinensis Object Detection

Abstract

In precision agriculture, object detection is often limited by the scarcity of image data, as agricultural data collection is inherently constrained by season, geographical location and crop growth cycles. Generative AI offers a practical way to expand limited datasets and enrich data diversity. However, generation artifacts may introduce unreliable supervision. To address the challenge of scarce image data, we propose ReSynDet, a Generate–Assess–Train framework for reliable use of synthetic images in data-scarce Brassica chinensis object detection. In the Generate stage, we introduce a Flux.1-fill-based out-painting pipeline to produce diverse synthetic images from individual real images without additional model training. In the Assess stage, we propose Degradation-Chain Ranking-based Generated Image Quality Assessment (DCR-GIQA), which learns image-level fidelity scores from ordered degradation chains and replaces manual quality annotation with self-supervised ranking. In the Train stage, we propose GIQA-guided Semi-Supervised Object Detection (SSOD), where synthetic images serve as unlabeled data in teacher–student SSOD without requiring additional manual annotation, and GIQA-derived scores are incorporated as image-level weights into the unsupervised loss to emphasize high-fidelity synthetic images while softly suppressing unreliable ones. Experiments with only 6 labeled real images containing 75 plant instances and 120 unlabeled synthetic images show that ReSynDet improves AP50:95 from 61.30 to 66.44 and F1-score from 71.86 to 76.99. These results support the effectiveness of ReSynDet in improving object detection under the evaluated data-scarce Brassica chinensis setting.