SDD-Net: self-supervised dual-domain dual-path single-pixel imaging

Abstract

Existing supervised deep-learning single-pixel imaging methods mostly require paired label data to pre-train the network. Such training methods consume a considerable amount of time to annotate the dataset and train the network. Additionally, the generalization ability of the network model limits the practical application of deep learning single-pixel imaging. Especially for complex scenes or specific applications, precise imaging details pose challenges to existing single-pixel imaging methods. To address this, this paper proposes a self-supervised dual-domain dual-path single-pixel imaging method. Using a self-supervised approach, the entire network training only requires measuring the light intensity signal values and projection pattern images, without the need for actual labels to reconstruct the target image. The dual-domain constraint between the measurement domain and the image domain can better guide the uniqueness of image reconstruction. The structure-texture dual-path guides the network to recover the specificity of image structure information and texture information. Experimental results demonstrate that this method can not only reconstruct detailed information of complex images but also reconstruct high-fidelity images from low sampling rate measurements. Compared with the current state-of-the-art traditional and deep learning methods, this method exhibits excellent performance in both imaging quality and efficiency. When the sampling rate is 5.45%, the PSNR and SSIM indicators are improved by 5.3dB and 0.23, respectively. The promotion of this technology will contribute to the application of single-pixel imaging in military and real-time imaging fields.