BCNN: An Effective Multifocus Image fusion Method Based on the Hierarchical Bayesian and Convolutional Neural Networks

Abstract

Because the focus information is obtained under different optical depth, it is impossible to collect all relevant information of objects from the only one image. The multifocus image fusion technique enables it to gather all of the focus data from the partially focused images, enhancing contrast and sharpness. To overcome the troubling weakness of the already-existing fusion methods, such as the incomplete boundary information and partial loss of focus, a new network called “BCNN”, combining the layered Bayesian and the convolutional neural network (CNN for short), is constructed. The hierarchical Bayesian can well maintain the texture features and edge information, and change the traditional way of learning a fixed value of the weight by learning the obvious features that are represented by the mean and variance. Meanwhile, the activity levels and the fusion rules can be jointly and deeply learned by the CNN model, avoiding the sophisticated plan and special design for the fusion rules. According to the aforementioned concepts, a novel BCNN-based fusion model for multifocus images is proposed. After detailed experimental implementation, the accuracy and efficacy of the proposed method are extensively illustrated and proved, not only in the way of the numeric evaluation, but also the highlighted visual comparison.