A method of duck egg surface defect detection based on improved GoogLeNet （GoogLe Net-Mini） was proposed, and the other three neural networks include GoogLeNet, VGG16 and AlexNet were compared. The results showed that the accuracy of the four networks were 95.88%, 94.16%, 92.75% and 85.43% respectively. The detection accuracy of GoogLeNet-Mini for three kinds of duck eggs （normal, dirty and damaged） was 98.43%, 97.45% and 95.88% respectively. Compared with GoogLeNet, VGG16 and AlexNet, GoogLeNet-Mini had higher accuracy, better generalization and robustness, and the detection accuracy of three types of duck eggs can meet the production requirements. The detection range is applicable to duck eggs with more than 5% dirty area and more than 2% damaged area.