Image denoising for single-pixel imaging based on cycle generative adversarial networks with attention mechanisms

Single-pixel imaging is an image processing technology based on mathematical models and computer algorithms. However, it often suffered significant image quality loss due to noise in the environment, especially when the measurement conditions were single-photon detection, under-sampling, and complex. Therefore, we provided a denoising method by integrating an efficient attention-incorporated CycleGAN denoising method (EA-CycleGAN) to address this issue. The efficient attention module can enhance the feature extraction ability of the networks and is suitable for a high-noise environment. The results of this paper show that the denoising effect of integrating CycleGAN with the efficient attention module (EA-CycleGAN) on the target image is better than that of the conventional CycleGAN method and is far better than the Gaussian filtering, nonlocal means, and BM3D methods. It confirms that the proposed method can significantly improve single-pixel imaging quality while exhibiting strong generalization capability.