<p>This paper presents a lightweight zero-reference deep learning approach for simultaneous low-light image enhancement and noise reduction. The proposed method addresses critical challenges in real world applications by developing an efficient three-module architecture: block partition-and-concatenation (BPC), image enhancement module (IEM), and noise reduction module (NRM). The BPC module enables processing of high-resolution images by dividing them into manageable sub-blocks, reducing memory usage and computational complexity. The IEM employs MPDiDCE-Net, a novel lightweight network incorporating multi-path structure, dilated convolution, and channel shuffle operations to estimate transformation parameter maps without requiring paired training data. The NRM utilizes a self-supervised learning framework based on depthwise separable convolution to effectively suppress noise amplification for visual quality improvement. Comprehensive evaluation on LOL and SICE datasets demonstrates superior performance compared to some existing methods. Quantitative results show that the proposed scheme achieves 2.6% and 0.33% improvements in PSNR and SSIM respectively over Zero-DCE, while reducing model size by 34.96%, FLOPs by 48.39%, and runtime by 22.22%. Qualitative assessment reveals enhanced visual quality with improved brightness and contrast while maintaining natural color reproduction. The results demonstrate that the proposed scheme can not only raise the visual quality but also decrease the model size. Moreover, experimental results demonstrate that the proposed scheme outperforms some existing zero-reference-based methods in terms of PSNR, SSIM, model size, FLOPs, and run time.</p>