Spectroscopy and Spectral Analysis, Volume. 45, Issue 2, 558(2025)
Improved Fusion Algorithm for Infrared and Visible Images Based on Image Enhancement and Convolutional Sparse Representation
Infrared and visible light images have become important source images in the field of image fusion research due to their complementary characteristics, and the current infrared and visible light image fusion methods have the problem of limited ability to preserve the details of texture information in the image. In this paper, firstly, the histogram equalization (HE) method is used to dynamically expand the range of gray values of infrared and visible images after alignment to achieve image enhancement, which makes the texture information in the image more prominent, and at the same time, the contrast between the background of the image and the details of the texture is also improved. Secondly, the gradient minimization filter is used to smooth the enhanced image to obtain the background layer of the image, and then the source image and the background layer are used to obtain the detail layer by difference operation to realize the decomposition of the infrared and visible light images. Again, the convolutional sparse representation (CSR) is combined with feature similarity analysis for infrared and visible image fusion: the two detail layers containing rich texture information are fused using the fusion strategy based on the convolutional sparse representation, and in this process to reduce the mismatch sensitivity of the convolutional sparse representation method, a window-based averaging strategy is adopted to process the activity level map of the image, to make the convolutional sparse representation insensitive to mismatches; For the problem of large amount of redundant information in the background image, the feature similarity analysis of the two background layers is carried out, and this is used as the basis for determining the degree of importance of the two in the fusion process. Finally, the preliminary fused detail and background layers are reconstructed by the inverse transform of gradient minimization image decomposition, and the fusion results of infrared and visible light images are finally obtained. Two sets of images, scenes 1 (buildings) and 2 (woods), from 21 scenes in the TNO dataset are used for subjective visual analysis. The observation results show that the HE-CSR-based fusion method visually retains the image's texture details better than the existing eight typical image fusion methods, including CVT, DTCWT, FPDE, GTF, IFEVIP, LP, RP, and CSR. At the same time, the objective index evaluation of the image fusion effect of all scenes in the TNO dataset is further conducted. The results show that the six evaluation index values of SF, SD, SCD, AG, EN, and CC for the HE-CSR-based fusion results are 7.316 6, 37.350 5, 1.704 1, 5.571 4, 6.756 3, and 0.744 6, which are respectively improved by 19.54%, 21.87%, 13.11%, 31.31%, 2.17%, and 8.23%. The experimental results show that the HE-CSR fusion method proposed in this paper outperforms other typical methods in subjective visual analysis and objective index evaluation and provides a new and more effective model and method for infrared and visible image fusion.
Get Citation
Copy Citation Text
ZHU Rong, ZHENG Wan-bo, WANG Yao, TAN Chun-lin. Improved Fusion Algorithm for Infrared and Visible Images Based on Image Enhancement and Convolutional Sparse Representation[J]. Spectroscopy and Spectral Analysis, 2025, 45(2): 558
Received: Oct. 30, 2023
Accepted: May. 21, 2025
Published Online: May. 21, 2025
The Author Email: ZHENG Wan-bo (zwanbo2001@163.com)