Under the impact of typhoon disasters, the judicious utilization of tie lines and emergency resources can enhance the resilience of distribution systems. Leveraging Deep Reinforcement Learning (DRL) technology, a two-stage resilience enhancement (TSRE) method for distribution systems is proposed to support model-free solutions to complex optimization decision problems, significantly achieving real-time dynamic decision-making. Firstly, the TSRE process is decomposed into pre-disaster flexible allocation of emergency resources and in-disaster coordinated dispatching of tie lines and emergency power vehicles. Secondly, a Markov decision process (MDP) is formulated, and a DRE simulation environment is established based on OpenDSS. The Deep Q Network (DQN) algorithm is adopted to solve the optimal policies for resilience enhancement. Finally, simulation experiments are conducted on the IEEE 33-bus distribution system. The experimental results indicate that compared with no-action strategy, the integrated resilience indexes are improved by 54.84% with an average reaction time of only 0.0016 s against disturbances.