This paper proposes a new feature selection algorithm. First, the data at every attribute are sorted. The continuously distributed data with the same class labels are grouped into runs. The runs whose length is greater than a given threshold are selected as “valid” runs, which enclose the instances separable from the other classes. Second, we count how many runs cover every instance and check how the covering number changes once eliminate a feature. Then, we delete the feature that has the least impact on the covering cases for all instances. We compare our method with ReliefF and a method based on mutual information. Evaluation was performed on 3 image databases. Experimental results show that the proposed method outperformed the other two.