Spectral analysis systems based on speckle patterns have become a research hotspot due to advances in the theory of spectral reconstruction. Research in this area focuses on optimizing scattering media and spectral reconstruction algorithms, as well as improving the cost-effectiveness of system integration solutions. The core performance of these systems depends on the algorithms’ ability to efficiently analyze speckle patterns, and the introduction of deep learning algorithms has significantly improved spectral resolution, bandwidth, and system robustness. However, when applying the aforementioned research approaches to infrared spectral detection scenarios, the speckle pattern acquisition process relies entirely on infrared indium gallium arsenide (InGaAs) cameras. Compared to more mature silicon-based cameras, such devices suffer from issues like insufficient sensitivity and high cost, limiting their practical application and hindering large-scale production. Therefore, developing an alternative technical solution that converts infrared light to be tested into the detectable wavelength range of silicon-based cameras holds significant practical importance. This approach is expected to further expand the application boundaries of speckle spectroscopy analysis systems across various fields.

An infrared spectroscopy detection scheme based on second-harmonic generation (SHG) and deep learning algorithms was proposed. This approach utilized a periodically poled lithium niobate waveguide to convert the infrared light to be detected into a wavelength range detectable by a silicon-based camera, and employed frosted glass—which is compact, offers excellent scattering properties, and is cost-effective—as the scattering medium to construct the spectroscopy detection system (Fig.1). Furthermore, by combining deep learning algorithms for broadband spectral reconstruction, the system ultimately achieved high-resolution spectral detection at a low cost.

Using the ResNet-50 residual neural network for single-wavelength identification, the model learned the nonlinear mapping relationship between wavelengths and their corresponding speckle patterns. This spectral analysis system achieved an accuracy rate of 99% or higher in correctly identifying speckle patterns with wavelength intervals of 0.1 pm in the 1549 nm, 1550 nm, and 1551 nm bands. The t-SNE clustering diagram visually demonstrated the identification results across each wavelength band. Speckle patterns corresponding to the same wavelength cluster into distinct groups, while clusters corresponding to different wavelengths were completely separated, forming 10 mutually exclusive and non-overlapping categories (Fig.2). Additionally, by incorporating deep convolutional neural networks (DnCNN) for denoising, the model effectively removed noise from the transmission matrix-reconstructed spectra. The system achieved sparse and broadband spectral reconstruction with a resolution of 0.1 pm across a 40 pm wavelength range (Fig.4). By comparing reconstruction errors with conventional truncated singular value decomposition (TSVD) algorithms, the proposed DnCNN model demonstrated excellent noise reduction performance. In particular, when processing sparse spectra, its reconstruction quality was improved by approximately 16 times compared with traditional algorithms. Based on the overall comparison of three reconstruction errors for sparse spectra, the DnCNN model achieved denoising performance that was one order of magnitude higher than conventional methods, successfully enhancing the system’s spectral resolution to 0.1 pm (Table 1).

An infrared speckle spectroscopy system based on a periodically poled lithium niobate waveguide is designed. By integrating nonlinear frequency conversion effects into speckle spectroscopy, this system focuses on the need for high-resolution infrared spectral detection. It enables compact, low-cost, high-precision, and user-friendly infrared speckle spectroscopy, which is of great significance for overcoming bottlenecks in infrared spectroscopy and advancing the transition of speckle-reconstruction spectrometers from laboratory research to industrial applications.