Abstract: Aiming at the problem of consensus enhancement in opinion dynamics of social network, a consensus opinion enhancement with intelligent perception (COEIP) model based on multi-agent reinforcement learning is proposed. In the Markov decision-making process in opinion dynamics, firstly, the decision-making model of agent is designed through bidirectional recurrent neural network to solve the problem of uncertain-length perception. Then, through the idea of difference reward, an effective reward function is designed for the three objectives of convergence efficiency, connectivity and communication cost. Finally, in order to optimize COEIP model, a multi-agent exploration and collaborative update algorithm based on policy gradient is designed, which can enable agents to adaptively learn the neighborhood selection strategy with multi-objective trade-off ability through the reward value in the process of interaction with each other. A large number of simulations verify that COEIP can effectively reconcile the contradictions between agents and reduce the number of clusters when the system is stable in the scenario of opinion dynamics of social network, thus enhancing the consensus opinion of the system. This model provides a new solution for improving the unity of people's opinions under large-scale social networks, which has important theoretical guiding significance.