AT-Net: A Semi-Supervised Framework for Asparagus Pathogenic Spore Detection under Complex Backgrounds

Asparagus stem blight is a devastating crop disease, and the early detection of its pathogenic spores is essential for effective disease control and prevention. However, spore detection is still hindered by complex backgrounds, small target sizes, and high annotation costs, which limit its practical application and widespread adoption. To address these issues, a semi-supervised spore detection framework is proposed for use under complex background conditions. Firstly, a difficulty perception scoring function is designed to quantify the detection difficulty of each image region. For regions with higher difficulty scores, a masking strategy is applied, while the remaining regions are adversarial augmentation is applied to encourage the model to learn from challenging areas more effectively. Secondly, a Gaussian Mixture Model is employed to dynamically adjust the allocation threshold for pseudo-labels, thereby reducing the influence of unreliable supervision signals and enhancing the stability of semi-supervised learning. Finally, the Wasserstein distance is introduced for object localization refinement, offering a more robust positioning approach. Experimental results demonstrate that the proposed framework achieves 88.9% mAP50 and 60.7% mAP50–95, surpassing the baseline method by 4.2% and 4.6%, respectively, using only 10% of labeled data. In comparison with other state-of-the-art semi-supervised detection models, the proposed method exhibits superior detection accuracy and robustness. In conclusion, the framework not only offers an efficient and reliable solution for plant pathogen spore detection but also provides strong algorithmic support for real-time spore detection and early disease warning systems, with significant engineering application potential.