Multi-focus image fusion via multi-scale attention and Siamese networks

Abstract

Multi-focus Image Fusion (MFIF) technology aims to generate a full-focus image with an extended focus range by combining multiple images with different focal depths. This has significant implications in fields such as image restoration and medical imaging. This paper proposes a new MFIF method based on deep learning, which utilizes multi-scale attention and a Siamese network structure to efficiently extract local depth features from images and enhance the fusion effect. The design of the Siamese network structure allows the model to process paired multi-focus images and share the feature extraction process in the deeper layers of the network. This not only enhances the expressive capability but also improves the model's ability to recognize images with different focal depths. Consequently, the network can effectively capture local depth features, which provides rich information for subsequent fusion. By incorporating a multi-scale dilated convolution attention module, which dynamically adapts the receptive field size to encompass a larger number of pixels, the process of information aggregation is facilitated across a wider area, thereby enhancing the optimization of the feature reconstruction process. Furthermore, binary segmentation and small-area filtering methods are employed to enhance the consistency of the fused image. Experimental results show that the proposed method surpasses existing multi-focus image fusion methods in terms of both subjective visual effects and objective evaluation metrics.