Single-pixel imaging is an innovative computational imaging technology that relies solely on illumination patterns and a single-point pixel detector with no spatial resolution to acquire target light intensity information. Subsequently, the reconstruction of the target image is achieved by computing the correlation between the intensity information and the corresponding illumination patterns. In comparison to traditional imaging technologies employing multi-point pixel array detectors, single-pixel imaging offers advantages such as a broad wavelength range of applications, minimal redundancy in data storage, and a simple system structure. Particularly, it exhibits superior imaging performance in certain challenging environments (e.g., dense foggy forests). This technology has garnered widespread attention in various crucial domains, including terahertz imaging [[1], [2], [3]], remote sensing imaging [4], three-dimensional imaging [[5], [6], [7]], optical computing [8,9], and other non-imaging optical signal processing areas [[10], [11], [12], [13]].

Previous studies have focused extensively on imaging methods, yielding numerous research outcomes such as Computational Ghost Imaging (CGI) [[14], [15], [16], [17]], Orthogonal Matching Pursuit Imaging (OMPI) [18,19], Total Variational Single Pixel Imaging (TVSPI), etc. [20]. These methods reconstruct approximate original images from compressed sampling values, which contain less information than the target scene. They offer novel research directions for imaging technology. However, these methods rely on a high number of sampling times, making it challenging to achieve clear reconstruction of target images within a short timeframe, particularly in the absence of expensive high-speed spatial light modulation equipment [21]. In recent years, owing to the widespread application of efficient deep learning methods [22,23]. The introduction of learning-based approaches has become a new trend in the single-pixel imaging domain, such as the Deep Learning-based Computational Ghost Imaging (DLCGI) method [24,25]. These methods combine traditional CGI with deep learning, and under medium-low sampling rates (10.00%–30.00%), they initially use CGI to generate blurred images, which are then gradually refined through network model calculations. Compared to CGI, OMPI, and TVSPI, etc., DLCGI can better address the mutual constraints between “sampling times” and “reconstruction image clarity” (i.e., the lower the sampling times in traditional single-pixel imaging methods, the poorer the clarity of the reconstructed image). However, DLCGI relies heavily on extracting feature information from blurred images. In actual test environments where disturbance information cannot be avoided, the clarity of CGI reconstruction from low-sampling conditions is further reduced. Consequently, the network models struggle to produce visually appealing reconstructed images and may encounter issues generating images of incorrect categories. Recognizing the limitations of DLCGI, some researchers have proposed end-to-end single-pixel imaging methods based on deep learning (End-to-end SPI) [[26], [27], [28], [29]]. These methods directly utilize less light-intensity information to reconstruct images, offering a new research direction for achieving high-quality image reconstruction under low sampling rates (≤10.00%). However, existing End-to-end SPI methods face several challenges: 1. The variation in sampling times: Under the same sampling rate, the number of sampling times significantly increases with the rising demand for reconstructed image resolution (i.e., assuming a sampling rate of 10.00%, an image requiring a reconstruction resolution of 32 × 32 pixels need 102 sample values, whereas an image requiring a reconstruction resolution of 128 × 128 pixels need 1638 sample values). 2. Singular imaging resolution: The reconstructed image resolution scales up at a 1:1 ratio with the resolution of the illumination pattern (i.e., projecting the illumination patterns with a resolution of 32 × 32 blocks onto the target will only generate a reconstructed image with a resolution of 32 × 32 pixels). 3. Poor anti-interference ability: In actual environments where both light scattering loss and system noise are unavoidable, enhancing illumination pattern resolution and increasing sampling times may introduce more disturbance, severely impacting the clarity of reconstructed images. Specifically, focusing on the actual applications of single-pixel technology, challenge 3 is deemed most crucial. Among them, the influence of enhancing illumination pattern resolution is illustrated in Fig. 1 which we have provided. (Note: research efforts are currently underway or have been made by scholars to reduce sampling times. Therefore, our emphasis leans towards the less-explored aspect of “illumination pattern resolution”). In Fig. 1(a), the sampling times are set to 9830 (i.e., corresponding to a reconstruction image resolution of 128 × 128 pixels with a sampling rate of 60%), utilizing the TVAL3 reconstruction algorithm. Asdepicted in Fig. 1(a), under noise-free conditions, higher-resolution illumination patterns projected onto the ground truth yield better imaging results. However, introducing 10% random salt-and-pepper noise into the ground truth introduces more disturbance when using higher-resolution illumination patterns, consequently affecting the clarity of the reconstructed image. In contrast, when employing low-resolution illumination patterns (e.g., 8 × 8 and 16 × 16 blocks) for single-pixel imaging, the variations in a and g (b and h) are minimal, and the diagonal pixel value curves of the reconstructed image tend to overlap. However, a challenge arises with the issue of imaging blurriness. Further, in the study of single-pixel imaging under disturbance environments, W. Gong in 2022 compared the active projection-based and passive encoding-based single-pixel imaging methods under light disturbance environments. He concluded that monitoring and correcting the reflection signal of the target before modulation during the image reconstruction process can improve the quality of passive encoding-based single-pixel imaging [30]. In 2022, Z. He et al. addressed the problem of poor suppression capabilities in single-pixel imaging network models and proposed a single-pixel imaging network structure based on a random dropout mechanism. This improved image quality from the perspective of enhancing the generalization capability of the network model [29]. In 2023, Q. Guan et al. considered that sampling values carrying noise represent one of the fundamental problems leading to the decline in imaging quality. They proposed a method to separate, recognize, and reduce noise from sampling values before reconstructing images [31]. Previous research contributed significantly to improving image quality from the perspectives of sampling values, network models, and imaging methods. However, studies focusing on the resolution of illumination patterns are still relatively uncommon.

Based on the discussion above, considering the strong anti-interference capability of low-resolution illumination patterns, this paper proposes a single-pixel imaging method via low-resolution illumination patterns. This method involves using a small number of low-resolution illumination patterns to compressively sample the target, acquiring rough light signals contains a small amount of disturbance. These signals are then processed through a network model to directly generate a high-resolution image. The research initially focuses on designing the structure of the single-pixel imaging network based on compressed sampling and deep learning. Subsequently, an experimental platform for the single-pixel imaging system is constructed. Finally, the comparative analysis is conducted between OMP [19], TVAL3 [20], DLBOGI [26], DSPINet [29], and the proposed CSSPI in both simulation experiments and actual tests.