Infrared imaging is an important means of modern battlefield reconnaissance, and target recognition technology based on infrared images can provide important support for intelligence interpretation. Aiming at target recognition in infrared images, a method based on selected deep features was proposed. A ResNet with a proper structure was designed to perform feature learning on infrared images, and the output feature maps from each convolutional layer was vectorized to obtain a corresponding feature vector. For the deep feature vectors of different feature maps, their correlations with the original image were evaluated based on the Spearman rank correlation coefficient. Afterwards, several deep features with high correlations were selected through the threshold decision algorithm. The deep features obtained after selection can eliminate unnecessary redundant components, thereby improving the accuracy and robustness of subsequent classification. The joint sparse representation model was used to characterize and classify the selected deep features, and finally the category of the sample can be identified. Therefore, the proposed method can effectively combine the discrimination of the multi-level deep features learned from ResNet, thereby improving the final recognition performance. The experiments were carried out in the public mid-wave infrared target image dataset (MWIR), using the original test samples, simulated noisy samples and simulated occluded samples to test and analyze the performance of the method. The experimental results show that the proposed method can achieve stronger effectiveness and robustness compared with some existing infrared target recognition ones.