China is a big country in the production of duck eggs and ducks. The duck egg hatching industry has a huge output. It needs to incubate billions of ducklings every year to meet the production needs. At present, the method of removing the infertile eggs in the duck egg hatching industry in China is to visually recognize the eggs by artificially photographing the eggs about 7 days after hatching. This method is inefficient and has no edible value after 7 days of incubation, which will cause huge waste of resources. Machine vision technology was used to hatch the third day of the duck eggs. The end-to-end characteristics of the deep convolutional neural network was used to the image of the duck egg on the third day of incubation, and it was directly input into the neural network, and the Alexnet neural network. The convolutional layer was used to replace the fully connected layer, and the size of the convolution kernel was changed. An egg net fertilization information recognition network (Eggnet) model was established to realize the nondestructive discrimination of the fertilization information in the early hatching of the duck eggs. The test results showed that the accuracy rate of the method for the classification of the duck eggs in the third day of hatching was as high as 98.87%, the accuracy of the verification set was 97.97%, and the average single egg detection time was only 0.24s. This technology can be used in the actual production of duck egg hatching industry in the later stage. It would replace the artificial egg method to select the infertile egg. It can solve the problem of automatic device installation in the egg hatching industry. It had broad application space.