Spectral reflectance reconstruction aims to calculate the true reflectance distribution of an object's surface in response to incident light from hyperspectral images. It is an important research topic in quantitative remote sensing and other hyperspectral imaging applications. Traditional reflectance reconstruction methods mainly rely on statistical models, often requiring stable flat-field illumination conditions or reference images as prior information for reflectance estimation. However, when reference images or other prior information are lacking in practical applications, especially under complex and varying lighting conditions, reconstruction accuracy using traditional methods is not ideal. This paper treats illumination estimation and reflectance reconstruction as a constrained matrix decomposition problem to address this issue. It proposes a multi-level cyclic optimization network model for spectral reflectance reconstruction. The model utilizes a hybrid channel-spatial attention mechanism to adaptively focus on key features in reflectance spectral images, thereby enhancing the extraction and amplification of critical information under non-uniform and multi-illuminant conditions, significantly improving reflectance reconstruction's robustness. Additionally, the network integrates a denoising mechanism comprising two modules: low-rank regularization and total variation regularization. The low-rank regularization module explores the intrinsic low-dimensional structures of illumination and reflectance to suppress noise interference. The total variation regularization module imposes spatial smoothness constraints on the reconstructed spectra, thereby improving reconstruction accuracy, reducing spectral mutations and redundant information, and ensuring spatial consistency throughout the process. To validate the effectiveness of the proposed method, this paper designs related data preprocessing, model training, and evaluation methods. The KAUST hyperspectral dataset is used as the training set in the training process, and different types of incident light source scenarios are simulated in the testing phase. Using the CIE 1964 10° standard observer color matching function as a reference, the hyperspectral images are converted into color images for visualization and quantitative performance evaluation. Experimental results show that the proposed reference-free reflectance reconstruction method outperforms traditional statistical-based reconstruction methods and current popular deep learning-based reconstruction methods regarding reconstruction accuracy indicators such as SAM and GFC. Particularly, without a calibrated whiteboard as a reference, the method still maintains high spectral reconstruction accuracy, demonstrating superior generalization capability and excellent reconstruction performance in complex lighting environments.