Abstract: The low-temperature deposition (LDM) food 3D printing process is easily affected by a combination of multiple factors, such as extrusion temperature, speed, and material rheological properties, leading to various issues such as blockage, excessive extrusion, and interference, affecting the printing quality. To achieve efficient identification of defects, this study developed a visual monitoring platform based on industrial cameras by constructing a multi-category nozzle extrusion defect image dataset, addressing the problems of small defect sizes and insignificant inter-class differences. A dual-channel multi-scale feature extraction network that integrated ResNet and Swin Transformer was proposed, and a multi-scale attention fusion module was introduced to improve recognition accuracy. Experimental results showed that the proposed model could effectively identify various types of extrusion defects, with prediction accuracies of 98.20%, 97.89%, 94.90%, and 97.86% for blockage, excess, granularity, and interference, respectively, and 97.63% for standard extrusion; the average detection time for a single image was 0.2 s. The overall accuracy was 97.53%, and the recall rate was 97.62%, which were 4.11 and 4.00 percentage points higher than those of ResNet34, respectively. This demonstrates excellent recognition performance and real-time capabilities, providing reliable data support for food 3D printing quality control.