Abstract: To address the severe scarcity of appearance defect samples in industrial high-speed packaging lines and the blind spots in single-view detection of three-dimensional packaging, a distributed visual perception and knowledge-driven decision fusion detection method for small-sample 3D packaging is proposed. Firstly, to solve the overfitting problem in training with small sample datasets, a dynamic online data augmentation strategy based on physical constraints is proposed. Through geometric perturbation and illumination perturbation, the generalization ability of the model is significantly improved without increasing the annotation cost. Secondly, to address the difficulty in feature alignment caused by asynchronous data collection from multiple positions, a distributed visual perception and knowledge-driven decision-making collaborative architecture is designed. Independent lightweight YOLOv8n detection models are used to extract local features, and a defect-view visibility map constructed based on process prior knowledge is combined with the knowledge-driven multi-view decision fusion strategy to achieve logical complementation of multi-view information. Finally, experiments are conducted using a typical soft box packaging product as the research object. The results show that the method effectively overcomes the problem of missed detection of concealed defects such as side warping and folding skewing with only a small number of original samples. The mAP50 reaches 94.9%, and it has the engineering application value of low cost and easy deployment.