DEDNet: An Infrared and Visible Image Fusion Architecture based on Dual-Encoder and Channel-Picking/Gaussian-Filtering Models

DenseFuse is a new approach for infrared and visible image fusion. Considering the single encoding strategy of DenseFuse, we propose a dual-encoder DenseNet (DEDNet), which develops a heterogenous image dual-encoder and a channel picking/Gaussian filtering based fusion strategy. The proposed method includes encoding layers, fusion strategy and decoder, in which encoding layers consist of dual encoders. Since infrared and visible images have different imaging mechanisms, the dual encoders can extract the features of infrared and visible images more effectively and improve the quality of fused images. The fusion strategy based on l1-norm's channel selection and Gaussian filtering improves the structural integrity and spatial correlation of the fused features. In the DEDNet, the infrared image is input to the infrared encoder to get the infrared features while the visible image is input to the visible encoder to get the visible features. Then, the fusion strategy fuses the infrared and visible features structurally and spatially. Finally, the decoder reconstructs the fused features to obtain the fused image. Experiments show that the DEDNet achieves competitive results in both subjective and objective evaluation metrics compared with other fusion approaches.