The work aims to propose a vacuum packaging defect detection method for pickled vegetables based on YOLOv5s network to solve the low efficiency and high leakage rate of manual-based vacuum defect packaging rejection of pickled vegetables. Firstly, Ghost Convolution was used to replace the convolution in the CSP module, which reduced the number of parameters in the network while improving the feature extraction capability of the model; Secondly, in order to reduce the loss of feature information in down sampling, the space-to-depth (SPD) and depthwise-separable convolution (DSConv) were used in down sampling; Finally, the SE attention mechanism module was introduced in the network to improve the accuracy of the algorithm. On the dataset of homemade pickled vegetable packaging, the mean average precision (AmAP) of the improved network reached 93.88 and the model size reached 3.91 MB. Compared with the original model, the mAP was increased by 2.05% and the model was reduced by 44.38%. The method in the paper enables the classification and localization of the defective vacuum packages of pickled vegetables, and lays a foundation for robot-based defective package rejection.