Infrared and visible image fusion based on dilated residual attention network

In recent years, deep learning (DL)-based techniques have achieved significant improvements over image fusion applications. Yet, current DL-based approaches raise formidable feature extraction, computational and statistical challenges in image fusion models. To overcome these challenges, we proposed an end-to-end DL-based architecture for infrared (IR) and visible (VIS) image fusion. We introduce multi-scale feature extraction and self-attention-based new feature fusion strategy to generate a high-quality fused image having balance details of IR and VIS modalities. Specifically, instead of using normal convolutions, we introduce dilated convolutions in the encoders to extract multi-scale features of IR and VIS images. Additionally, we introduce self-attention mechanism to refine and adaptively fuse multi-contextual features of IR and VIS images. Fused image is generated via decoder of the network. Extensive qualitative and quantitative evaluations on a benchmark dataset illustrate that our proposed method achieves reasonable performance over other state-of-the-art and current CNN-based image fusion methods.