摘要:近年来，卷积神经网络凭借极强的学习能力，在图像复原任务上实现了比传统学习方法更令人满意的结果.但是，由于丢失了重要的纹理细节，这些基于卷积神经网络的方法普遍存在着复原图像过度平滑的缺点.为解决该问题，提出一种基于级联密集型卷积神经网络的轮廓波域图像复原方法，可以应用于单幅图像去噪、超分辨率及JPEG解压缩这3个经典图像复原任务.首先，构建了一种紧凑的级联密集型网络结构，不但可以充分挖掘和利用不同层次的图像特征，而且解决了由于网络加深带来的长期依赖问题.接着，引入可以稀疏表示图像重要特征的轮廓波变换，分别将低质量图像和重建图像对应的轮廓波子带作为网络的输入和输出，更加有效地恢复出逼真的结构和纹理细节.在标准测试集的实验表明：提出的方法在3个图像复原任务上达到了当前最优的性能，不但获得了更高的峰值信噪比和结构相似度，而且在主观的重建图像中包含了更加真实的纹理细节.



Abstract:In recent years, due to the powerful learning ability, convolutional neural networks (CNN) have achieved more satisfactory results than conventional learning methods in image restoration tasks. However, these CNN-based methods generally have the disadvantage of producing over-smoothed restored image due to the fact that losing important textural details. In order to solve this problem, this study proposes an image restoration method based on cascaded dense CNN (CDCNN) in contourlet transform, which can be used for three classical image restoration tasks, namely, single image denoising, super resolution, and JPEG decompression. First, this study constructs a compact cascading dense network structure, which not only fully exploits and utilizes the different hierarchical features of images, but also solves the problem of the long-term dependency problem as growing the network depth. Next, this study introduces the contourlet transform into CDCNN, which can sparsely represent the important image features. Here, the contourlet subbands of low-quality image and corresponding restored image are used as the input and output of the network respectively, which can recover realistic structure and texture details more effectively. Comprehensive experiments on the standard benchmarks show that the unanimous superiority of the proposed method on all three tasks over the state-of-the-art methods. The proposed method not only obtains higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), but also contains more realistic textural details in the subjective reconstruct images.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5866