A new multi-focus image fusion quality assessment method with convolutional sparse representation

Assessing image fusion quality purposefully is a challenging task due to the diversities of fused features. In this work, a specific multi-focus image fusion quality assessment method is proposed based on joint image layering and convolutional sparse representation. Specifically, the proposed method includes two stages: Tikhonov regularization optimization-based joint image layering and convolutional sparse representation-based focus similarity comparison. The first stage decomposes the source images and their fusion result jointly into a common base layer and respective detail layers, and then, the second stage compares the focus similarity between these detail layers with their convolutional sparse features. The main novelty of our work is to assess fusion quality with learning features, rather than with those handcrafted low-level patterns. Consequently, our method has higher reliability and feature-level analytical ability. A large number of objective and subjective experiments demonstrate the effectiveness and specificity of the proposed method. Moreover, the applicability of the general blind natural image quality metrics for image fusion was also examined and discussed. Besides experiments, the feature-level characteristics of multi-focus image fusion were also investigated and analyzed with the proposed method. Our analysis reveals some potential laws that could provide new perspectives for fusion algorithm design and improvement.