Research On Pixel-level Multi-focus Image Fusion Algorithms

Multi-focus image fusion is one of information synthesis methods,mainly used to fuse multiple source images that are focused on different regions from the same target scene.The application of multi-focus image fusion can improve the utilization of image information and ensure the reliability and accuracy of the subsequent processing of the system,so it is widely used in computer vision,medical imaging,military and etc.Pixel-level multi-focus image fusion is the basis of multi-focus image fusion due to the high retention of original information and detail features.Correctly distinguishing between focused and out-of,focused pixels in the source image is the key to pixel-level multi-focus image fusion.Generally,spatial domain based fusion methods use a single feature as a pixel clarity metric,which leads to the poor robustness of pixel classification.The method proposed in this paper is to construct a convolution neural network and use its "learning" ability to obtain a new kind of pixel clarity metric.The main contents of this thesis are as follows:(1)A multi-focus image fusion algorithm based on convolution neural network for block classification is proposed.This algorithm is derived from the fusion method based on image block.For the measure of block clarity,a new way is use a convolution neural network to classify clear and fuzzy block and get the classification score as a clarity metric.Firstly,creating a training set contained a large number of clear and fuzzy block samples by artificially.Then,adjusting the structure of CNN step by step for the purpose of improving classification accuracy.Finally,the CNN for block classification will be trained by the training set produced before.The image fusion experiment shows that the method applied the block classification CNN to the multi-focus image fusion is better than the genera block based image fusion methods.(2)A multi-focus image fusion algorithm based on convolution neural network for image segmentation is proposed.Block based image fusion method generates a large number of overlapping blocks,which resulting in high computational cost.Therefore,an improved method with image segmentation convolution neural network is proposed.The new CNN can classify every pixel in source image directly,without splitting the image into blocks in advance.Consequently,using this network for image fusion can greatly reduce the cost of computing.The fusion experiment shows that the improved algorithm has a significant improvement over the computational efficiency of the previous algorithm,and the quality of the fusion image is improved accordingly.Finally,the main contents of this paper are summarized,and the further work can be put forward.