<p>Infrared-visible image fusion plays a crucial role in object detection and recognition under complex environmental conditions. Traditional fusion algorithms often fail to achieve a balanced trade-off between infrared target enhancement and visible texture preservation, leading to compromised overall perceptual quality. To address this challenge, this study proposes a novel fusion network (FusionNet++), which incorporates explicit attention and residual refinement mechanisms, building upon previous quantitative analyses of conventional fusion architectures. The proposed network adopts a dual-branch encoder to extract multi-level features from infrared and visible input images, while an explicit attention mask adaptively assigns pixel-level fusion weights to emphasize salient regions in the infrared spectrum. Additionally, a residual refinement module with amplitude constraints is integrated to suppress fusion artifacts and recover intricate structural details. The model is trained in an unsupervised manner using a hybrid loss function focused on infrared saliency, which includes luminance preservation, structural similarity (SSIM), gradient consistency, and total variation constraints. Experimental results demonstrate that the proposed algorithm improves the visibility of small infrared targets while preserving visible details, outperforming traditional methods and several advanced deep learning-based fusion approaches in terms of entropy (EN), gradient preservation, SSIM, and probability of detection (Pd). This study provides a novel structural perspective and practical insights for multispectral image fusion and the detection of tiny infrared targets.</p>