To solve the undue sensitivity of Newton iteration to the selection of initial value and the disorder of the features extracted by using ICA (Independent component analysis), an improved algorithm based on fast ICA and optimum selection is proposed for infrared (IR) objects classification. The algorithm conducts one dimension search on the direction of Newton iteration to ensure convergence of search results. Meanwhile, a novel rule is designed for selecting effective classification features according to distance function. Thus declining of classification rate and robustness with the increasing of training samples is eliminated. Experimental results demonstrate that the proposed algorithm can provide higher classification rate with fewer object features and is more robust in different kinds of classes compared with the traditional methods.