Research On Multi-focus Image Fusion And Quality Evaluation Method Based On Multi-classification Model

Focus classification learning based method is the most popular framework method for multi-focus image fusion.This type of spatial domain approach treats multi-focus image fusion as a focus/defocus classification problem: first,a pre-trained neural network model is used to determine the focus properties of the source image,and then an all-in-focus synthetic image is generated by stitching the partial fused source images.Although this deep learning-based fusion framework has shown outstanding performance,the fusion quality is degraded due to the fact that the basic premise of focused binary classification of source images is not fully realistic.In order to make full use of the advantages of deep learning models and effectively overcome these shortcomings,this paper proposes a multi-classification focus model and applies it to the design of fusion algorithms.The degradation of the focus/defocus boundary caused by the simple binary classification model can be effectively overcome and the overall fusion quality can be significantly improved by describing the regional states in the source image as in-focus,out-of-focus and uncertain and performing targeted processing.In addition,based on the idea of focused classification learning,this paper also introduces deep learning into image fusion quality evaluation for the first time,and proposes a new objective evaluation index for multi-focus image fusion.This method has better applicability and accuracy than traditional evaluation indicators.The main work of this paper includes:(1)A multi-focus image fusion algorithm based on multi-classification focus model is proposed.The algorithm distinguishes three different categories of pixels according to the multi-classification focus model,and designs fusion rules to fuse them separately,which can effectively overcome the problem of poor focus/defocus boundary quality caused by the two-class model.Embedding multi-scale decomposition can to a certain extent overcome the lack of robustness of the single-scale stitching rules commonly used in focused classification learning methods for unregistered source images.The specific steps include obtaining a focus probability map by training a CNN classifier to classify the source image and reconstructing the probability values,decomposing the source image and the focus probability map by using Laplacian pyramid and Gaussian pyramid,and obtaining the three-classification focus property map(TCFPM)according to the decomposed focus probability map and the multi-class focusing model at each level,executing the multi-scale fusion strategy according to the TCFPM,and inversely transforming the Laplacian pyramid to obtain the fusion image.(2)This paper proposes a multi-focus image fusion quality evaluation method based on the multi-classification focus model.First,a high-performance classifier is used to obtain the focus probability map,and then it is converted into a focus property map TCFPM according to the multi-class focus model.Then the corresponding similarity comparison rules are designed to compare the similarity between the fusion image and the source image in units of image patches.Finally,a intuitive scoring graph and overall quality score are obtained.Scientific experiments have proved that the proposed evaluation index is correct and effective.