There exist a large number of irrelevant and redundant features in large-scale infrared spectrum datasets. To solve this problem, a dynamic weighted infrared spectrum feature selection algorithm (MBDWFS) is proposed. The algorithm deletes the irrelevant and redundant features in an original spectrum dataset by combining the symmetric uncertainty metrics with Markov Blanket. Then, a smaller scale optimal feature subset is obtained. By comparison with three classical feature selection algorithms FCBF, ID$_3$ and ReliefF, it shows that the proposed MBDWFS algorithm is better than the above three algorithms in overall classification performance and is more suitable to be used in the field of material infrared spectrum analysis.