Attention-Guided Ensemble CNN for Multi-Focus Image Fusion

Abstract

The Convolutional Neural Networks (CNNs) based multi-focus image fusion methods have recently attracted enormous attention. They greatly enhanced the constructed decision map compared with the previous state-of-the-art methods that have been done in the spatial and transform domains. Nevertheless, these methods have not reached satisfactory initial decision maps, and they require extensive post-processing algorithms to achieve acceptable results. In this paper, we propose the Attention-Guided Ensemble Convolutional Neural Network, which extends the Ensemble CNN (ECNN) framework by incorporating channel-wise attention via Squeeze-and-Excitation (SE) blocks and cross-path gradient attention mechanisms. These attention modules adaptively recalibrate feature representations, improving focus discrimination and interpretability without requiring post-processing. The proposed method achieves superior performance on benchmark datasets including the Lytro dataset and individual test images, with only a ~4.9% parameter increase over the ECNN baseline. Extensive experiments demonstrate consistent improvements of 1.85-2.56% across multiple evaluation metrics compared to state-of-the-art methods, while maintaining computational efficiency. The attention mechanisms enable the network to focus on informative channels and effectively integrate gradient information from multiple paths, resulting in cleaner decision maps and sharper fused images.