Multi-scale geometric analysis is a popular tool that is widely used in the field of multi-focus image fusion. It plays a large role in extracting the features from input images. In this paper, a novel multi-focus image fusion method based on a non-fixed-base dictionary and multi-measure optimization is presented in the non-subsampled shearlet transform (NSST) domain. The proposed approach contains the following four steps. The input images are first decomposed by NSST into low- and high-frequency coefficients. Second, a sparse representation (SR)-based framework is proposed and applied to merge the low-frequency coefficients of the input images. In this framework, a non-fixed-base dictionary iteratively trained by the previous dictionary is constructed, which represents the complex details of the source images. Third, a type-2 fuzzy logic scheme is introduced for the high-frequency coefficient fusion, which can effectively select the high-quality coefficients from the source images so that it can achieve better performance. Finally, an inverse NSST operation is conducted in the merged low- and high-pass subbands and thus obtains the initial fused image. A multi-measure optimization method is then employed to optimize the initial fused image, and thus, the final fused result is achieved. In this method, three different measures, namely, pixel difference, visual saliency, and similarity, are designed to select the focus regions more completely. The experimental results demonstrate that our proposed approach yields a better effect than other methods in both the visual quality and the objective assessment.