A growing number of complex features are designed to address various problems in computer vision. Feature selection is an efficient way to reduce the heavy computation cost caused by complex and lengthy features. The total features are substituted by a discriminative subset selected according to a criterion to reduce the dimensionality. In current feature selection methods, most of the subset selection metrics evaluate the features according to their relevance. However, the discriminative power of a feature subset is not simply determined by the relevance, especially in the case of complex features. In this paper, a new feature subset selection metric, the image-to-class distance ratio based on Euclidean distance, is proposed to select a subset in which the average intra-class Euclidean distance is minimized and average inter-class Euclidean distance is maximized, leading to good classification performance. In addition, the search space for feature selection problems with large-size complex features is huge, which makes many heuristic search methods infeasible. A Particle Swarm Optimization (PSO) based subset search algorithm is introduced to search the best subset according to the proposed metric in the huge search space. Experimental results show that the proposed subset search algorithm (I2CDRPSO) is effective. The feature subset selected by the metric has better performance than other feature selection metric in several classification methods.