Compared to high-resolution spectra, low-resolution spectra offer higher observational efficiency and broader sky coverage, making them especially valuable for large-scale stellar surveys. The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) survey alone has collected tens of millions of low-resolution stellar spectra, providing an unprecedented opportunity for large-scale stellar parameter estimation. However, a substantial portion of these spectra suffer from low signal-to-noise ratio (low-SNR), which poses significant challenges for accurate parameter determination. Accurately extracting stellar atmospheric parameters from such data can significantly enhance the utility of spectral observations. However, these low-SNR spectra often introduce considerable uncertainty in parameter estimation. To address this issue, we propose a novel method based on the Cycle-Consistent Convolutional Neural Network (Cycle-CNN) for predicting key stellar atmospheric parameters, including effective temperature (Teff), surface gravity (log g), and metallicity ([Fe/H]). This method integrates the cycle-consistency learning mechanism of Cycle-GAN with the strong modeling capability of CNNs, thereby improving model robustness and reducing prediction uncertainty under low-SNR conditions. We train and evaluate the model on spectra from LAMOST DR9 across different SNR intervals (2–15). For spectra with SNR between 10 and 15, the model achieves prediction accuracies of 63.22 K for Teff, 0.11 dex for log g, and 0.07 dex for [Fe/H]. For the spectra with SNR between 5 and 10, the prediction accuracies are 89.45 K, 0.17 dex, and 0.11 dex, respectively. Even under extreme conditions with SNR between 2 and 5 and limited data availability, the model maintains good performance, achieving accuracies of 145.36 K, 0.29 dex, and 0.18 dex. Furthermore, we validate our predictions against reference parameters from high-resolution surveys, and the results demonstrate good consistency with other large-scale spectroscopic surveys. These findings indicate that the proposed Cycle-CNN method can provide stable and accurate predictions of atmospheric parameters even under low-quality spectral conditions, offering a reliable solution to improve the scientific utilization of low-quality spectra.