Joint supervised and unsupervised deep learning method for single-pixel imaging

To solve the problem of low image quality that under-sampled single-pixel imaging (SPI) often suffers from, deep learning based SPI methods have attracted more attention recently. However, the image reconstruction quality is apt to be restricted due to the limitation of network structures in capturing long-range dependencies. Moreover, deep learning based methods show a significant performance degradation when modulation patterns change slightly. In this paper, we propose an effective method based on channel attention convolutional neural network for under-sampled SPI. The method can reconstruct high-quality object images directly from SPI measurements and guarantee the strong generalization ability by taking advantage of unsupervised deep learning. Meanwhile, it effectively avoids over-fitting problem using SPI model constraint and total variation regularization. Extensive experimental results on simulation and real data demonstrate that the proposed method has superior performance in image quality, noise robustness and generalization compared with the state-of-the-art SPI methods.

Automated summary

In this paper, we propose an effective method based on channel attention convolutional neural network for under-sampled single-pixel imaging (SPI) to reconstruct high-quality object images directly from SPI measurements. The method takes advantage of unsupervised deep learning and effectively avoids over-fitting problem using SPI model constraint and total variation regularization. Extensive experimental results on simulation and real data demonstrate that the proposed method has superior performance in image quality, noise robustness, and generalization compared with the state-of-the-art SPI methods.