基于子空间投影的生成对抗网络红外图像降噪

Infrared Image Denoising Based on A Generative Adversarial Network Improved by Subspace Projection

摘要:

针对红外图像信噪比低、边缘信息模糊、杂波干扰多等检测难点,提出一种基于子空间投影的生成对抗网络红外图像降噪方法。首先,生成器由U-Net结构和子空间注意力网络构成,编码阶段由4层下采样实现图像特征提取,解码阶段由4层上采样重建图像。其次,在每层跳跃连接中加入子空间投影网络,每层特征图与同层上采样后图像共同放入子空间投影网络进行图像特征融合,将投影特征图与原始高级特征融合实现图像降噪。最后,将图像输入到鉴别器进行对抗训练,得到清晰重建图像。实验结果表明,与BM3D,DnCNN等常用算法相比,改进的生成对抗网络算法有更好的客观评价指标效果,PSNR和SSIM分别达到了34.36,0.9852dB,从而验证了改进算法的良好降噪性能。

Abstract:

To address the detection challenges of infrared images, such as a low signal-to-noise ratio, blurred edge information, and clutter interference, a generative adversarial network infrared image denoising method based on subspace projection is proposed. First, the generator consists of a U-Net structure and a subspace attention network. The encoding stage extracts image features through four layers of downsampling, while the decoding stage reconstructs the image through four layers of upsampling. Second, a subspace projection network is added to each skip connection, and the feature maps of each layer are combined with upsampled images from the same layer to form a subspace projection network for image feature fusion. The projected feature maps are then fused with the original high-level features to achieve image denoising. Finally, the image is input to the discriminator for adversarial training to obtain a clear reconstructed image. The comparative experiments with BM3D(Block-Matching and 3D Filtering), DnCNN(Deep Neural Neural Network For Image Denoising), and other algorithms show that the improved generative adversarial network algorithm has better objective evaluation index effects, with PSNR and SSIM reaching 34.36dB and 0.9852, respectively, thus verifying the strong denoising performance of this algorithm.