Due to limitation of optical lenses, obtaining all-in-focus images is difficult. However, lots of multi-focus image fusion methods cause undesirable artifacts around the focused and defocused boundaries in fusion images. Usually, these boundaries are at the edges of objects in images while the gradient information can reflect edge information intuitively. Based on the above ideas, a Gradient-based method using convolution neural network (CNN) is proposed to produce all-in-focus image. Specifically, we transmit the original images and corresponding four kinds of gradient images into five CNN models to generate the five initial focus score maps, respectively. Then, the final segmented focus map is obtained via merging the initial focus score maps. Finally, we combine the final segmented focus map and source images to obtain the fused image. The experimental results demonstrate that the proposed method has a better performance on both quality and quantitative evaluations than other state-of-the-art methods.

由于光学透镜的限制，难以获得全焦点图像。但是，许多多焦点图像融合方法会导致融合图像中聚焦和散焦边界周围出现不良伪像。通常，这些边界位于图像中对象的边缘，而渐变信息可以直观地反映边缘信息。基于上述思想，提出了一种基于卷积神经网络（CNN）的基于梯度的方法来产生全聚焦图像。具体来说，我们将原始图像和相应的四种梯度图像传输到五个CNN模型中，以分别生成五个初始聚焦得分图。然后，通过合并初始聚焦得分图获得最终的分段聚焦图。最后，我们将最终的分割后的聚焦图和源图像结合起来以获得融合图像。