Ro-FusionGAN:An Adversarial Framework for High-Quality Multi-focus image fusion

Multi-focus image fusion (MFIF) aims to synthesize an all-in-focus image from source images with varying focal depths. While deep learning has advanced this field, existing "process-imitation" approaches often rely on binary mask supervision, leading to boundary artifacts and a dependency on heuristic post-processing. To address these limitations, we propose Result-Oriented Fusion Generative Adversarial Network (Ro-FusionGAN), a novel result-oriented adversarial framework. Unlike methods that mimic intermediate focus maps, our framework directly optimizes the perceptual quality of the final fused image via a composite fusion-aware loss function. Furthermore, we introduce a differentiable Total Variation (TV) regularizer to autonomously enforce spatial smoothness, enabling the generation of soft decision maps and eliminating the need for post-processing. Extensive experiments demonstrate that Ro-FusionGAN outperforms eleven state-of-the-art methods in visual quality, quantitative metrics, and computational efficiency, yielding artifact-free images with natural focus transitions.