| Due to the limited depth of field of the camera,it is difficult to obtain a fully focused image for the same scene.The most important reason behind this is that objects at a specific distance from the lens are in focus and sharp,while other objects may be blurred.In order to extend the freedom of the lens and improve the quality of acquired images for more practical applications,a low-cost image processing technique is needed to make each object a fully focused image.This can be achieved by integrating multiple images with different focus planes,so that the resulting composite image is a fused image in which all objects are in focus.This technique of image processing is called multi-focus image fusion(MFIF).As an important branch in the field of image fusion,MFIF has been a hot research topic.Focus region detection is a key element of MFIF technology.In this thesis,we focus on MFIF based on focus region detection.Based on the three most commonly used paths in image fusion,i.e.,spatial domain,transform domain and deep learning,three different MFIF methods are proposed to effectively solve the challenging problems encountered in the current MFIF process,such as block effect,artifacts,and distortion and low fusion efficiency.The main research work of this thesis is described as follows.1.A new optimal block-based MFIF method is proposed to address the traditional fixed block-based fusion method which is prone to block effects and other problems in the fusion results.Different from the traditional block-based method,the proposed method in this thesis does not use fixed blocks,but adopts a quadtree decomposition to process the source images.First,in order to detect the focus information of the source image,a new edge-weighted SML based on the Sum of Modified Laplacian(SML)is proposed as the focus measure,named SEWML.This improved focus measure is more robust compared to SML.Then,an effective quadtree decomposition strategy is proposed to decompose the source image into optimally sized blocks.Meanwhile,SEWML is used to detect the focused blocks in the quadtree structure of the source image,naturally combine the focused blocks to form the initial decision map,and optimize the decision map to obtain the final decision map.Finally,the fused image is obtained by the weighted average rule according to the final decision map.Experimental results on different types of multi-focused source image pairs datasets show that the proposed method in this thesis obtains better fusion performance compared with 14 other state-of-the-art methods.2.A MFIF method based on linear sparse representation and image keying is proposed to address the problem that the existing MFIF methods do not capture the focus/defocus boundary(FDB)information accurately.First,a focus measure based on linear sparse representation is introduced,which uses the linear relationship between the dictionary formed by the natural image and the input image on the local window to represent the focus information of the image by solving for the linear coefficients.Then,this focus measure is used to obtain the focus map of the source image and a ternary map consisting of the focus region,the scattered focus region,and the unknown region containing the FDB,and the ternary map is used as an input to process the FDB region of the source image using image keying techniques to obtain a more accurate fully focused image.Finally,in order to further improve the quality of the fused image,the obtained full-focus image is used as a new dictionary to realize the fusion process iteratively,and the final full-focus fused image is obtained after a set number of updates.The image fusion experiment uses a typical30-pair multi-focused source image.The experimental results show that the proposed method has better fusion performance and visual effect with higher computational efficiency compared with other state-of-the-art fusion methods.3.A new MFIF method based on deep regression learning is proposed to address the problem that existing deep learning-based image fusion methods commonly use chunks as input.The method is an end-to-end structure and treats focal region detection as a regression problem.Specifically,a newα-matte boundary scattering model is first employed to accurately model the scattering diffusion effect and generate more realistic training data for multi-focus images.Then,a pair of multi-focused source images is fed directly into the network to predict their associated binary masks,where each element indicates whether the input pixel is focused or scattered.Thanks to this pixel-to-pixel regression learning,the distinguishing information between focused and scattered pixels and the complementary information between each pair of pixels are effectively utilized,thus solving the boundary ambiguity problem present in other chunkingbased methods.Finally,the fused images are generated by a weighted averaging rule based on the estimated masks.The method in this thesis further improves the quality of the fused image by introducing the edge retention measure and the focus information measure.Extensive experimental results show that the method proposed in this thesis achieves better results in both visual perception and quantitative analysis compared with other state-of-the-art fusion methods. |
