Abstract: Estimating the direction of arrival （DOA） of a given object in environments with a low signal-to-noise ratio （SNR） remains a challenging problem in array signal processing. The covariance matrix contains information about the spatial correlation of signals and the distribution of noise and can thus be regarded as a special type of “image”. Both image denoising and low-SNR DOA estimation share the core task of extracting useful structured information from contaminated data. Although existing deep learning methods such as convolutional neural network （CNN） models can extract spatial features from the covariance matrix， they often lose fine details in conditions with strong noise interference， which leads to a notable degradation in performance. To address these challenges in estimating DOA in environments with low SNR， a novel deep learning network based on a CNN-NAFNet cascaded architecture is proposed. The proposed approach was inspired by advances in image denoising， especially the nonlinear activation-free network （NAFNet）， and adopts a three-stage strategy including coarse feature extraction， refined enhancement， and global aggregation. First， a multi-scale CNN is applied to perform coarse-grained feature extraction from the input covariance matrix. Multi-scale features are aggregated from local details to global context to form a robust preliminary representation through progressive downsampling and channel expansion. Subsequently， these coarse features are refined using an NAFNet enhancement module designed for the purpose， which suppresses noise interference by leveraging SimpleGate and simplified channel attention （SCA） mechanisms. Multiple learnable residual scaling factors are incorporated to dynamically adjust the contribution of residual connections and mitigate gradient vanishing. Finally， global feature fusion is performed and a fully connected classifier outputs the probability distribution of target angles to provide an end-to-end estimation of the DOA of a given object or target. The simulation results demonstrate that the proposed method achieved superior estimation performance under low-SNR conditions compared to existing approaches. Furthermore， they also show that the method exhibited excellent robustness and generalization capability while meeting the real-time requirements of practical array signal processing systems.