摘要
泛锐化方法通过融合低分辨率高光谱图像和全色图像,从而获取具有高空间、高光谱分辨率的图像。为充分挖掘全色图像空间纹理特征,在VGG19中引入八度卷积,同时关注全色图像不同空间尺度的高低频特征;为增强光谱图像的局部空间特征,添加特征调制模块,将全色图像高频特征注入光谱图像空间细节;为加强光谱图像特征同全色图像低频特征的全局依赖关系、降低网络复杂度,网络深层对多头注意力进行优化,使光谱图像只关注全色图像最相关的图像区域;重建阶段采用Sobel滤波器、GCN图神经网络以及ECA注意力聚合光谱图像边缘细节和显著波段特征,同时,引入VGG感知损失和边缘损失,提高融合后图像的边缘特征比重。在Pavia Center、Botswana和Chikusei这3个数据集与其他泛锐化方法进行验证,其中峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)值分别达到40.80%、45.90%和43.26%,SAM值分别降低到4.15%、1.38%和2.70%。提出泛锐化方法可有效恢复光谱图像空间纹理细节,同时避免光谱失真。
The pan sharpening method combines low resolution hyperspectral and panchromatic images to obtain images with high spatial and spectral resolution.To fully explore the spatial texture features of panchromatic images,octave convolution is introduced into the original VGG19 feature extractor,while paying attention to the high and low frequency features of panchromatic images at different spatial scales.To enhance the local spatial features of spectral images,a feature modulation module is added to the original VGG19 feature extractor to inject high-frequency features of panchromatic images into the spatial details of spectral images.In order to enhance the global dependency between spectral image features and low-frequency features of panchromatic images,and reduce network complexity,the network optimizes multi head attention deeply,so that spectral images only focus on the most relevant image block regions of panchromatic images.In the reconstruction stage,Sobel filters,GCN graph neural networks,and ECA attention aggregation spectral image edge details and significant band features are used.At the same time,VGG perception loss and edge loss are introduced to increase the proportion of edge features in the fused image.Verified with other pan sharpening methods on the Pavia Center,Botswana,and Chikusei datasets,the Peak Signal-to-Noise Ratio(PSNR)values reached 40.80%,45.90%and 43.26%,respectively,while the SAM values decreased to 4.15%,1.38%and 2.70%,respectively.The proposed pan sharpening method can effectively restore spatial texture details of spectral images while avoiding spectral distortion.
作者
郭伟
蒋鹤
王春艳
GUO Wei;JIANG He;WANG Chunyan(College of Software,Liaoning Technical University,Huludao 125105,China)
出处
《微电子学与计算机》
2024年第11期48-59,共12页
Microelectronics & Computer
基金
国家自然科学基金青年基金(41801368)。
关键词
高光谱图像
泛锐化
特征增强
高低频特征
hyperspectral image
pan sharpening
feature enhancement
high and low frequency feature
