Abstract: To solve the problems of limited infrared features and poor feature matching in the process of infrared image stitching, this paper proposes an algorithm called sparse depth feature infrared image stitching (SDFS). The algorithm first extracts a dense depth feature map using a convolutional neural network, then calculates and describes sparse feature points from the feature map to enhance the quality of feature point extraction. Next, the K-nearest neighbor search method is used to perform coarse matching of sparse feature points, followed by the application of a dynamic distance ratio strategy to refine the matching results and improve matching accuracy. Finally, based on the matching results, a homography matrix is calculated for image projection transformation, and adaptive factor-weighted fusion is used to achieve seamless fusion and splicing of the image. Experimental results show that the algorithm exhibits high robustness and can effectively adapt to infrared image stitching in different scenes. The stitching accuracy and display effect outperform commonly used stitching algorithms based on SIFT or SURF feature extraction.