For adults’ facial appearances changing slowly, the age estimation of adults in adjacent age groups is still a challenge. Aiming at this problem, this paper introduced the adversarial training method into the age estimation and proposed an age estimation method based on Adversarial Dropout Regularization(ADR). The age feature learner and the discriminator are trained via the adversarial training method, then the ability of age feature learning(especially the adjacent age groups features) gets improved. Experimental results on three classic datasets (UTKFace, MORPH and Adience) show that the proposed model improves the accuracy of UTKFace from 42.8% to 81.6%, and improves the accuracy of MORPH from 39.8% to 69.8%. Moreover, the accuracy of Adience is 63.3%. Being compared with other 4 classic models, the model in this paper using the neural networks of 5 layers achieves better results than other deep neural networks, and outperforms other methods with averagely 17.5% higher accuracy for the young and middle-aged(15-53 years old), which shows that our model improves the performance significantly on age estimation task, and has the practical value.