Multi-focus Image Fusion Based On Deep Learning

With the development and progress of imaging technology and the popularization of various imaging devices,people can easily obtain a large number of images with different types and rich contents.For many computer vision and image analysis tasks,an image in which all objects in the scene are clear can provide more useful information.However,under actual shooting conditions,it is almost impossible to obtain an all-in-focus image that satisfies the application requirements in only one shot.There are two main reasons for this: First,the optical lens in the imaging device have a limited depth of field.Only objects within a certain distance from the optical lens can get focused,and objects outside the range will be out of focus;The second is that in the real scene,there are often many objects of interest at different depths.Multi-focus image fusion technology is to fuse multiple images focused on different depths taken in the same scene through image processing to obtain an all-in-focus image.In the context of the above research,this paper conducts deep research on multi-focus image fusion based on deep learning technology,with the aim of improving the quality of fused images.The main research work of this paper is as follows:1.Analyzed the current mainstream multi-focus image fusion algorithms,and summarized them into three categories: fusion algorithms based on transform domain,fusion algorithms based on spatial domain and fusion algorithms based on deep learning.The processing steps of these three types of methods are listed,and the advantages and disadvantages of them are analyzed and summarized.2.For the deep learning-based fusion algorithm,a large number of labeled multi-focus images are required as training data.This article uses a synthetic method based on the VOC 2012 natural image dataset with object segmentation annotation.The point spread function simulates defocus blur and synthesizes a sufficient number of multi-focus image pairs as the training dataset.3.Aiming at the problem that traditional transform domain and spatial domain-based fusion algorithms need to be carefully designed for focus level measurement and fusion rules,this paper introduces deep learning technology to convert focus detection into binary classification.Through convolutional neural networks,the mapping relationship between the source images and the focus decision map is established,and the fusion process of the multi focus images is guided by the focus decision map generated by the network directly.At the same time,in view of the current multi-focus image fusion algorithms that cannot perform complete focus detection on homogeneous areas in the image and detect inaccuratly at the boundary between focus and defocus regions,a novel U-shaped convolutional neural network structure is proposed,and a hybrid loss function is designed to improve the performance of network.Experimental results show that the proposed network structure and loss function effectively solve the above difficulties.In this paper,a comparison experiment is conducted with the existing algorithm on the public multi-focus image test dataset.In order to make a comprehensive and multi-dimensional quantitative evaluation of the fusion results of different algorithms,this paper uses four types of image quality evaluation indicators based on information theory,image characteristics,structural information,and human perception to quantitatively analyze and compare the fusion images.This article also uses a visual subjective perception evaluation method for qualitative comparison.The quantitative and qualitative experimental results consistently illustrate our algorithm in this paper can accurately detect the focus region of the source image and obtain a high-quality all-in-focus image.