一种基于深度卷积自编码的全色锐化方法

A Pansharpening Method Based on Deep Convolutional Auto-Encode

由于多光谱(Multi-Spectral,MS)图像具有较高的频谱分辨率,全色(Panchromatic,PAN)图像具有较高的空间分辨率,对MS图像和PAN图像进行融合,得到频谱及空间分辨率均较高的融合图像,这个过程称为全色锐化。针对当前深度学习研究中的全色锐化过程,容易出现过拟合,以及泛化能力下降,出现信息丢失和网络性能不够的问题,文中提出了一种基于深度卷积自编码的全色锐化方法,采用HSV变换,通过对色调分量进行编码与解码等过程生成融合图像,使得融合后的图像得以增强,并结合卷积神经网络的优势,克服传统全色锐化方法中对细节和纹理捕捉不足、噪声敏感度高问题,能有效提取图像的局部和全局特征。

Since the MS image has high spectral resolution and the PAN image has high spatial resolution,attempts are made to fuse the MS image and the PAN image to obtain a fused image with high spectral and spatial resolution,a process known as Pansharpening.For the current Pansharpening process in deep learning research,it is prone to overfitting,as well as the generalization ability decreases,information loss and insufficient network performance.A deep Pansharpening method based on convolutional CAE is proposed in this paper,using the HSV transform,to generate fused images through the process of encoding and decoding the hue channel,which makes the fused image enhanced.Combine the advantages of convolutional neural network to overcome the problems of insufficient detail and texture capture and high noise sensitivity in traditional Pansharpening methods,and effectively extract the local and global features of the image.