基于CNN和融合目标的三通道小波滤波器组识别  

THE RECOGNITION OF THREE-CHANNEL WAVELET FILTER BANKS BASED ON CNN AND FUSION TARGET

作　　者：刘斌 李昕 Liu Bin;Li Xin(School of Computer and Information Engineering,Hubei University,Wuhan 430062,Hubei,China)

机构地区：[1]湖北大学计算机与信息工程学院,湖北武汉430062

出　　处：《计算机应用与软件》2024年第2期209-215,285,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61471160)。

摘　　要：为解决目前需要人工选取二维不可分小波滤波器实现图像融合的问题,提出一种基于CNN和融合图像清晰度的二维三通道不可分对称小波的滤波器组自动择优分类方法。构造大量分布均匀的3×5对称小波滤波器组,并用其对多聚焦图像进行融合,根据融合结果对滤波器组设置融合清晰度高低的标签,并构造滤波器组的训练集和测试集;设计出分类的卷积神经网络,并进行训练得到模型;对训练集和测试集以外的滤波器样本进行识别与结果分析。实验结果表明:所设计的网络模型在测试集和测试集以外的数据集上的识别率分别为99.48%和99.58%,其分类结果中较好的滤波器类对多聚焦图像融合都有较高的清晰度。In order to solve the problem of manual selection of two-dimensional non-separable wavelet filter for image fusion,this paper proposes an automatic selection and classification method for of three-channel non-separable symmetric wavelet filter banks based on CNN and fusion image definition.Lots of well-distributed 3×5symmetrical wavelet filter banks were constructed.The multi-focus images were fused by using these filter banks.And the labels of fusion definition were set according to the fusion results along with constructing the training set and testing set.A classified convolutional neural network was designed,and a new network model was acquired after training.Filter banks outside the training set and the testing set were recognized and analyzed.The experiment results show that the recognition rate of the network model to the data set that is inside or outside the testing set is 99.48%and 99.58%respectively,both of which have higher recognition rate,and the better class of the filter banks has higher definition for multi-focus image fusion.

关 键 词：多聚焦图像融合 二维三通道不可分小波 CNN 滤波器组 清晰度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]