Summary: We propose a novel algorithm based on deep convolutional neural network to gain the Pareto optimal front (POF) of multi-objective optimization problem (MOOP). We transform the problem of gaining POF to a problem of image processing. Some vectors are sampled from the decision space and the objective function values of them are calculated to generate training data sets in a form of image. Then two deep convolutional neural networks are trained on the training data sets to predict POF. We have tested this algorithm on nine classical problems of MOOP and compared two important metrics of POF with the results of NSGA-II [K. Deb, A. Pratap, S. Agarwal and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II”, IEEE Trans. Evolutionary Comput. 6, 182–197 (2002)]. It is indicated that the POF’s metrics of our algorithm are better than NSGA-II’s over five test problems, especially the most difficult test problems ZDT4 and ZDT6. And predicting a result through our algorithm occupies no more than 7.75ms when the networks have been well trained for the specified problems. Additionally, it also provides a new view of the combination of deep convolutional neural networks and MOOP.