Macroscopic aspergillus recognition using YOLO-CSM

Aspergillus detection is of great significance in medical diagnosis and microbiological analysis. However, due to the complex morphology of Aspergillus, its subtle structural characteristics, the presence of rare strains, and partial occlusion in samples, traditional detection methods often fail to achieve high accuracy. To address these challenges, this study incorporates three innovative modules—CGNet, Shape-IoU, and MSCA—into the YOLOv12 framework to construct an efficient Aspergillus detection model. CGNet enhances the fusion and representation of complex structural features; Shape-IoU introduces morphological constraints to improve matching accuracy in scenarios with occlusion and blurred boundaries; and MSCA strengthens the model’s adaptability to strains of different scales. Experimental results show that the proposed model achieves a Precision of 92.1%, Recall of 87.8%, and mAP of 89.2% in Aspergillus detection, while requiring only 28.6M parameters, effectively balancing accuracy and lightweight design. These results demonstrate the superiority of the proposed method.