In image-based applications, such as medical imaging, remote sensing, and surveillance, the high performance and robust image restoration techniques are required. Although convolutional neural networks (CNNs) have outperformed traditional methods, they often struggle with a trade-off between spatial precision and contextual awareness. This paper proposes a Low-Illumination Image Restoration Network (LIRNet), a novel Recursive-MultiScale Residual architecture designed to accommodate this trade-off and optimize image restoration, particularly in challenging conditions like night and low-light scenarios. LIRNet employs parallel multi-resolution convolution streams to simultaneously capture local details and global context, addressing the limitations of existing approaches. An information exchange mechanism enhances feature representation, while spatial and channel attention mechanisms dynamically focus on relevant features, improving overall image quality. Furthermore, an attention-based multi-scale feature aggregation module ensures the preservation of high-resolution details and strong contextual information in the final output. Extensive evaluations on LoL standard dataset prove LIRNet’s superior performance in various image restoration and enhancement tasks.