Aiming at the problems that the existing food packaging box surface defect detection methods have difficulty in small target defect detection under complex background, high missed detection rate and low detection accuracy, with the mung bean cake snack packaging box common in life as the detection object, the work aims to propose a mung bean cake packaging box surface defect detection method based on improved Faster R-CNN. Firstly, based on the Faster R-CNN algorithm architecture, Swin Transformer V2-T was used as the feature extraction backbone to preliminarily improve the ability of the algorithm to extract the features of the packaging box defect. Secondly, combined with the weighted bidirectional feature Pyramid Network (BiFPN), the weight of each scale feature map was adjusted and the multi-scale fusion was conducted on features of different sizes to improve the recognition accuracy. Finally, ROIAlign was combined with the ECA attention mechanism to replace ROIPooling, removing two quantization errors and further optimizing the detection ability of the algorithm for packaging box defects. The detection method proposed could accurately extract the target defects, and the Average Precision (AP) of the four defects on the surface of the mung bean cake packaging box increased by 19.66%, 12.96%, 14.56%, and 18.86% respectively. At the same time, the mean average precision (mAP) increased by 15.76% when the IOU was 0.5. The improved model provides useful reference and experience for the application of Faster R-CNN in the intelligent production of food packaging boxes.