Abstract:
Deep learning is used for infrared and visible image fusion. If only deep features are extracted without feature processing, the fusion performance will be degraded in some aspects. Aiming at this problem, a fusion framework based on deep features and Zero-phase Component Analysis (ZCA) is presented. Firstly, the source images are decomposed into the low-frequency and high-frequency parts, and then the low-frequency sub-band parts were fused using the saliency detection. Secondly, for the high-frequency sub-band parts, the deep features are extracted using the residual network (ResNet), and then the deep features are normalized through the ZCA and
l1-norm regularization, the initial weight maps are obtained, and that the final weight maps are obtained by employing a Softmax operation in association with the initial weight maps. Finally, the fused image is reconstructed using a weighted-averaging strategy. Compared with the existing fusion methods, experimental results demonstrate that the proposed algorithm achieves better performance in both objective assessment and visual quality.