Multi-focus Images Fusion Based On Deep Convolutional Neural Network

Due to the limitation of the depth of field(DOF)in image sensor,it is difficult to get the image which contains all objects focused.The main reason is that only the objects in the distance of DOF can be focused.In one image,the focused objects are sharper than the defocused objects.In order to get all objects that focused in one image,a common solution is multi-focus image fusion technology.It aims to integrate two or more than two source images,which taken from the same scene,into one image.So that all objects can be focused in the fused image.This fused image contains more details and is more suitable for human vision and computer processing.Therefore,multi-focus image fusion can also be regarded as a process to improve the quality of a group of images.To address the problems in multi-focus image fusion,this paper focuses on the following two aspects: 1)The deep convolutional neural network based multi-focus image fusion method 2)The Hessian matrix based multi-focus image fusion method.The main contents are as follows:Proposed a pixel-wised convolutional neural network(p-CNN),which can recognize the focused and defocused pixel in source images from its neighborhood information,for multi-focus image fusion.The proposed p-CNN can be thought as a learned focus measure(FM)and performs more efficiency than conventional handcrafted FMs.To enable p-CNN with the strong capability for discriminating focused and defocused pixels,a comprehensive training image set based on a public image database is created.Furthermore,by setting precise labels according to different focus level and adding various defocus masks,the p-CNN can accurate measure the focus level of each pixel in source images that the artifacts in the fused image can be efficiently avoided.We also propose a method to implement the p-CNN with conventional image convolutional neural network(image-wised CNN)which achieves almost 25 times faster than using p-CNN directly in multi-focus image fusion.Experimental results demonstrate that the proposed method is competitive with or even outperform the state-of-the-art methods in terms of both subjective visual perception and objective evaluation metrics.Proposed a Hessian matrix-based multi-focus image fusion method.In this method,multi-scale Hessian matrix is utilized to detect feature and background regions.On this basis,source images are split into two different parts,and different fusion strategies are applied to generate decision map respectively.By combining decision maps in different parts,an initial decision map is obtained,and then the initial decision map is refined with post-processing method.To improve the performance of the fusion method,we propose new focus measures based on multi-scale Hessian matrix for both feature and background regions.Integral images are also introduced for fast computation to meet the real-time application requirements.Experimental results demonstrate that the proposed method is competitive with or even outperform the state-of-the-art methods in terms of both subjective visual perception and objective evaluation metrics.