Abstract: Source number estimation faces significant challenges under low signal-to-noise ratio (SNR) and limited snapshot conditions. To address this issue, this paper propose a deep learning-based model that combines feature extraction and feature discrimination, referred to as the Multi-receptive Field Non-local Denoising Residual Convolutional Network (MFNL-DRCN), to improve source detection accuracy. Firstly, a Multi-receptive Field Non-local (MFNL) module is proposed, which leverages multi-scale receptive fields and a non-local mechanism for feature extraction, capturing multi-scale spatial features and global dependencies of the signal. Subsequently, to reduce the interference from noise during the feature extraction process, an Attention Denoising Module (AD) is proposed, which integrates attention mechanisms and a soft-thresholding function to enhance the expression of key information and improve feature discrimination capabilities. Experimental results demonstrate that, under low SNR and limited snapshot conditions, the proposed method outperforms other techniques in terms of source number estimation accuracy.