Crowdsourcing is a powerful data collecting paradigm with human intelligence. By solving problems that are difficult for computers (such as finding a same person in a group single person pictures [1, 2]), crowdsourcing has shown strong advantages in recent research. In this paradigm, the large complex task will be decomposed into microtasks then distributed to workers over the world through the Internet. Various data collecting approaches are used with crowd. In this work, we explore the problem crowdsourced top-k queries using pairwise comparisons. Since human workers will be paid for their efforts, to minimize the total monetary cost, we carry out microtask-level and query-level cost minimization effort to achieve the goal. In the microtask-level which means the comparison on each pair, we use Confidence Interval(CI) and SteinJudge model to do minimization cost on each pairwise comparison. Microtask cost is proportional to the task difficulty, our model progressively asks human workers according to answers collected such that it can finish publishing tasks once they reach the stop condition. While in query-level which is the number of pairs compared, we design a RandomSelect algorithm to minimize query cost. Finally, we apply both synthetic data and real data to our model and algorithm, as the experiment results show that our methodology performs well and is very close to the optimal.