Superpixel-level Structural Similarity Metric for Image Fusion Quality Assessment

It is highly desirable to carry out an objective metric to evaluate the fused image quality. Existing metrics always extract image local features from image rectangular blocks to achieve assessment. However, the fixed shape of image block does not fit the natural attributes of the images and it is also not consistent with the perceptual characteristics of human visual system (HVS). In this paper, we presented a super-pixel level structural similarity metric for fused image quality assessment. Firstly, image features including brightness, gradient amplitude and texture complexity are extracted from adaptive super-pixel regions rather than rectangular blocks. Then, structural similarity of super-pixel level features between fused image and source images are calculated. Finally, the final metric is obtained by weighting average all local metrics according to super-pixel saliences. Experimental verifications are carried out in infrared and visible images, multi-focus images and medical images. Moreover, the experiments on stability analysis are carried out on a set of infrared and visible image sequence. The experimental results show that the proposed quality evaluation method has better performance and it is closer to the human visual evaluation results.