Research On Multi-focus Image Fusion Algorithm

In recent years,with the application of a large number of imaging equipment in human life and industrial production,multi-focus image fusion plays an increasingly important role in image processing.Affected by optical lens range of depth of field,optical lens can access to a single target scenario clear images of the whole scene,need many times on different parts of collection,get more focus on different parts of the source image,integration of the source image into a clear images of the scene is more focusing on image fusion of the important problems need to be solved.In this paper,deep learning multi-focus image fusion algorithm based on neural network is studied in depth.The main research contents are as follows:1.Based on the principle and implementation process of multi-focus image fusion,the paper analyzes and studies the pixel-level image fusion,discusses the transformation domain and spatial domain methods respectively by using typical algorithms,and analyzes the image scenes applicable to various algorithms.The quality evaluation method of image fusion and the quality evaluation index used in this paper are studied to provide a theoretical basis for the comparison and comparison of the fusion performance of subsequent algorithms.2.Studied the neural network based on supervised learning more focus on image fusion algorithm,to build a supervised image fusion algorithm based on CNN model,using a full convolution neural network based on supervised learning image fusion algorithms,the algorithm will be more focus on image fusion as a pixel prediction problem,by monitoring the convolutional neural network learning training,make the network to the complementary relationship of the original image in different focus for study,so as to achieve a global image.Simulation experiments and data analysis show that the fusion performance of this algorithm is better than that of traditional algorithms in all aspects,but it also has disadvantages such as complex computation and long time.3.A deep learning fusion algorithm based on unsupervised learning is proposed to solve the problems of supervised learning,such as the need to spend a lot of time and energy to create data sets and the long training cycle.First,an unsupervised codec network is trained to extract the deep features of the input image,and then these features and spatial frequencies are used to measure the pixel activity of the image and obtain the decision graph.Finally,the consistency check method is used to adjust the decision graph to get the final decision graph.Through the experimental results and data analysis of 16 fusion methods,it is proved that this method achieves good fusion effect in both objective evaluation and subjective evaluation.