Mutil-focus Image Fusion Based On Fully Convolution Neural Network

When the optical lens images an object,the object distance of the target will be different due to the difference in the scene,and the external factors such as the light intensity and the shooting angle will be different,resulting in the different definition of each target in the formed image.Image fusion technology can solvethe problem of rapid increase in the amount of information due to the increase in multi-sensor applications,and the problem that a single sensor cannot focus on all targets at the same time.This technology makes full use of the source data captured by multiple sensors and integrates multiple images of the same scene.The included information is integrated into a composite image to obtain comprehensive information describing the target more clearly and comprehensively.Image fusion has been widely considered to be of great significance in the fields of medical auxiliary detection,traffic detection,and clear image reconstruction.Multi-focus image fusion is an important embranchment of image fusion.It focuses on image processing and has been widespread used in various fields.This article mainly does further research in the direction of multi-focus image fusion,and proposes corresponding improvement methods.This thesis mainly does further research in the direction of multi-focus image fusion and proposes corresponding improvement methods.Traditional multi-focus image fusion methods,such as spatial domain method and transform domain method,have simple fusion strategies,but they all need to manually set the two key factors of activity level measurement and fusion rules,which have strong subjectivity and uncertainty.Its limited image transformation and complex fusion rules,as well as the lack of effective image fusion image representation methods make it enter a bottleneck.The main research purpose of this thesis is to combine the deep learning algorithm to solve the problems of the above-mentioned traditional image fusion methods.This type of method uses focus evaluation based on learning methods to replace artificially set activity level measurement and fusion rules to generate multi-focus fusion images with more clear pixels,which has good clarity and visual sensory effects.However,for CNN In the process of downsampling and convolution,not only the edge information of the source image is lose,but the local features such as the color and texture of local objects also interfere with the global semantic information used to distinguish the focus and defocus regions.Hence,A multi-focus image fusion method based on global feature extraction U-Net is proposed.The main research contents of this thesis are as follows:1.In view of the basic concept of multi-focus image fusion,this thesis comprehensively discusses the relevant theories and algorithms.Firstly,from the limitations of the image sensor itself,it analyzes the principle of multi-focus image generation and fusion.Secondly,the proposed multi-focus image fusion methods are classified and illustrated,and their pros and cons are discussed.Finally,image to image translation,convolutional neural network(CNN),U-Net and other related theories are reviewed.2.The production process of the data set used in this thesis and the relevant operation process of data preprocessing are introduced in detail.This thesis first introduces the process of image fusion,and explains the importance of image preprocessing in the process of image fusion.Then it introduces the generation of data sets and the necessity of making high quality data sets.In this thesis,the production method of the data set is introduced in detail,and three methods of data set preprocessing are introduced,and the application scene is analyzed,as well as the importance of data preprocessing.3.In view of the convolution operation too much attention to the local area,and will focus on figure generated as local classification problem,in order to increase the global feature coding ability of the network,this thesis proposes a global feature extraction U-Net image fusion algorithm,the introduction of the characteristics of the global pyramid extraction module and global attention to the sampling module,in order to effectively extract and using global semantics and edge information,and add sensory loss to the loss function,built a large-scale data set in order to improve the performance of CNN.The experimental results show that the fusion performance of the proposed algorithm is superior to other multi-focus image fusion algorithms in subjective evaluation,objective evaluation and network complexity.