Multi-focus Image Fusion Based On Focus Region Detection

Digital imaging equipment uses optical elements such as lenses or sensor arrays to capture visible light reflected from objects and record information about the scene.However,because optical lenses are limited by depth of field,it is difficult for objects in the same scene to be focused at the same time,resulting in different sharpness of front and rear scenes.Multi-focus image fusion technology can effectively capture objects located in different depths,so that two or more partially focused images can be fused into fully focused images.The main contents of this thesis are as follows:1.In order to solve the problems of supervised learning,lack of natural data sets,and large amount of network computing,this paper proposes an unsupervised hybrid model called FD-Fuse based on fractal dimension(FD).The key to this model is to combine traditional image processing methods with deep learning,reducing the number of model parameters and saving computational power on hardware devices.In addition,the feature extraction network based on convolutional attention module also improves the final fusion effect.Firstly,FD-Fuse uses unsupervised training of encoder and decoder networks to obtain shallow and deep features of the input image.Then,the pixel level differential box counting method proposed in this article is used to calculate the activity level of depth features and obtain the initial decision map.Next,the consistency verification method is used to obtain the final decision map,and finally,the pixel weighted fusion rule is used to obtain the final image.The experimental results show that compared to the 9 existing multifocal image fusion methods,FD-Fuse achieves the most advanced fusion performance in indicator evaluation and visual evaluation.2.In order to solve the problem of poor fusion effect caused by fixed block in block-based multi-focus image fusion method,a multi-focus image fusion algorithm based on adaptive block and guide filter is proposed in this thesis.In order to achieve better image segmentation,a cyclic differential evolution algorithm is proposed to avoid the influence of hyperparameters on the final result.In addition,in order to improve the fusion efficiency,population similarity is taken as the termination condition to avoid invalid iteration of the algorithm.Firstly,the cyclic differential evolution algorithm is used to calculate the optimal block size of the input image.Then,the sharpness matrix of the source image is obtained by using the definition standard function.Finally,a guide filter is used to guide fusion,enriching the details of the junction between focusing and defocusing in the final decision graph.Experimental results show that the proposed algorithm has better performance than five existing multi-focus image fusion methods.