In this paper, we present an efficient deep learning architecture for real-world smartphone image denoising. The architecture is developed based on attention mechanism in a multi-patch hierarchical network with non-local Information. Unlike traditional methods that struggle with the spatially variant and complex noise patterns in smartphone images, our model integrates a multi-patch hierarchical approach to effectively leverage spatial context at multiple scales. The proposed network incorporates a non-local module in the encoder to capture long-range dependencies, enhancing the network’s ability to model global image structure. To further refine feature representation, we introduce a parallel attention mechanism in the decoder that combines both channel attention (CA) and pixel attention (PA). This dual attention design enables the network to emphasize relevant features both across channels and within local spatial regions, leading to improved denoising performance. Trained on real-world datasets, including spatially variant noise from smartphones, our method demonstrates superior quantitative and qualitative results while maintaining a lightweight architecture with fewer parameters. Experimental evaluations validate the effectiveness of our model.