Activity level measurement and fusion rule are crucial factors in image fusion. Existing fusion methods often employ local filters and elaborate rules to measure activity levels and compare clarity information of source images, resulting in a clarity/focus map. This map contains integrated clarity information, which is significant for various image fusion tasks such as multi-focus image fusion and multi-modal image fusion. However, achieving satisfactory fusion performance with these methods is challenging. This research addresses this challenge through a deep learning approach, aiming to learn a direct mapping between source images and the focus map. To accomplish this, a deep convolutional neural network (CNN) is employed, trained using high-quality image patches and their blurred versions to encode the mapping. The key innovation lies in jointly generating the activity level measurement and fusion rule through the learned CNN model, overcoming the difficulties faced by existing fusion methods. This paper primarily proposes a novel multi-focus image fusion method based on this idea. Experimental results demonstrate that the proposed method achieves state-of-the-art fusion performance in terms of visual quality and objective assessment. The computational speed of the proposed method using parallel computing is sufficiently fast for practical usage. Additionally, the potential of the learned CNN model for other image fusion tasks is briefly exhibited in the experiments.