基于注意力机制和感知损失的遥感图像去噪

Remote sensing image denoising based on attention mechanism and perceptual loss

摘要:

遥感图像去噪对于遥感图像在后续的分类、检测等任务中有着非常重要的作用.为了让去噪后的图像更好地保留边缘细节信息，同时增强深度网络对噪声污染区域的辨识能力，本文结合注意力机制以及感知损失来处理遥感图像，提出了一种新的基于残差自编码器的遥感图像去噪网络ARED-VGG.考虑到遥感图像中不同地物大小不同，该网络首先同时使用图像的空间和光谱信息来提取多尺度特征；然后使用残差自编码器网络结构来进行图像空间-光谱多尺度图像重建.为了增加网络的辨识能力，更多地关注网络中提取的高频特征，网络引入了注意力机制.同时为了让去噪后的结果更符合视觉感观，使用了感知损失混合均方误差作为损失函数.从实验结果知，本文所提出的方法在噪声去除和纹理细节保留方面与NLM3D、BM4D、LRMR、HSID-CNN和3DADCNN相比表现更优.在Washington DC mall遥感图像数据集上进行了仿真实验，平均峰值信噪比以及平均结构相似性指标都有较好的结果；在AVIRIS Indian Pines数据集上进行了真实数据实验，以去噪后的结果地物分类指标作为验证，整体分类精度以及Kappa系数分别为96.90%和0.9647；对网络结构进行了消融实验，在两个数据集下，本文所提出的网络结构都能获得更优的结果.本文提出基于注意力机制和感知损失的深度神经网络进行遥感图像去噪，提高了网络的辨识能力，实现了良好的去噪性能，并且有效保持了图像的细节信息和光谱信息.

Abstract:

Remote sensing image denoising plays a very important role in the subsequent classification and detection tasks for remote sensing images. In order to retain more the edge details information in the denoised images and enhance the discrimination for contaminated regions, in this paper, the authors propose a new Attention based Residual Encoder-Decoder network aided by VGG (ARED-VGG) for remote sensing image denoising, which combines attention mechanism, and perceptual loss. First, considering that the different ground objects have different sizes, the network utilizes both spatial and spectral information to extract multiscale features. Second, in order to better represent the extracted features and improve the denoising effect, the residual autoencoder network is used in image reconstruction stage. Moreover, convolution operations that are responsible to extract high-frequency features should be paid more attention to facilitating the location of noise-affected regions. Based on these considerations, the attention mechanism is introduced to adaptively modulate feature representation. Finally, perceptual loss is adopted to maintain the visual results consistent to the human perception. In the experiments, the proposed method demonstrates superior performance in noise suppression and detail preservation to NLM3D, BM4D, LRMR, HSID-CNN and 3DADCNN. Simulation experiments were carried out on the Washington DC mall remote sensing image data set, d the results of mean peak signal-to-noise ratio and mean structure similarity index of the proposed method are better than the other methods. Real experiments were carried out on the AVIRIS Indian pines data set. The classification results for the denoised images were evaluated. The overall classification accuracy and kappa coefficient were 96.90% and 0.9647, respectively. The ablation experiment was carried out on the network structure on both data sets the proposed network architecture achieved best results. This paper proposes a deep neural network based on attention mechanism and perceptual loss to denoise remote sensing images, which improves the recognition ability of the network, achieves good denoising performance, and effectively preserves the detailed information and spectral information of the image.