Due to the limitations of infrared optical diffraction and infrared detectors, the noise of infrared images is relatively large and the resolution is low. Super-resolution reconstruction of infrared images improves image resolution, but at the same time enhances the noise of background. Aiming at this problem, a salience region super-resolution reconstruction algorithm for infrared images based on sparse coding is proposed. By combining the saliency detection and the super-segment reconstruction, it improves the target definition and reduces the background noise. Firstly, image feature is extracted by double-layer convolution, and image patches with large entropy are adaptively selected for training the joint dictionary. Sparse features are used to calculate the saliency to obtain salient regions, which reconstructs image patches in saliency region by the trained dictionary while the background region adopts Gaussian filtering. Experimental results show that the improved reconstruction algorithm is better than ScSR and SRCNN under the same conditions. The image signal-to-noise ratio is increased by 3-4 times.