关键词: 图像压缩感知, 深度学习, 卷积神经网络, 自注意力, 图像重构, 迁移学习

Abstract:

The existing optimization-inspired networks for image compressive sensing(ICS) implement information optimization and flow in the pixel domain following the traditional algorithms, which does not make full use of the information in the image feature maps extracted by the convolutional neural network. This paper proposes the idea of constructing information flow in the feature domain. A feature-space optimization-inspired network(FSOINet) is designed to implement this idea. Considering the small receptive field of the convolution operation, this paper introduces the self-attention module into FSOINet to efficiently utilize the non-local self-similarity of images to further improve the reconstruction quality, which is named FSOINet⁺. In addition, this paper proposes a training strategy that applies transfer learning to the ICS reconstruction network training for different sampling rates to improve the network learning efficiency and reconstruction quality. Experimental results show that the proposed method is superior to the existing state-of-the-art ICS methods in peak signal to noise ratio(PSNR), structural similarity index measure(SSIM) and the visual effect. Compared with OPINENet⁺ on the Set11 dataset, FSOINet and FSOINet⁺ have an average PSNR improvement of 1.04dB/1.27dB respectively.

Key words: image compressive sensing, deep learning, convolutional neural networks, self-attention, image reconstruction, transfer learning