<p>Efficient Image Super-Resolution (EISR) is critical for deployment on resource-constrained devices, requiring a balance between reconstruction fidelity and computational efficiency. Lightweight networks often face limited feature capacity, leading to detail loss and artifacts. We propose the Denoised Semantic-Guided Residual Network (DSRN), a compact framework that integrates semantic guidance and targeted artifact suppression. The backbone employs residual blocks with dilated convolutions, enlarging the receptive field for richer feature extraction. A multi-branch denoising module with dilated convolutions and channel attention suppresses checkerboard artifacts and anomalous textures while capturing long-range dependencies. Semantic guidance via pseudo-label fusion and a lightweight, randomly initialized feature extractor for perceptual loss further enhances texture reconstruction. A Segmented Training Strategy progressively optimizes individual components before joint fine-tuning, improving convergence stability. With structural re-parameterization for efficient inference, DSRN achieves state-of-the-art performance among lightweight SR models, particularly on complex textures, while remaining highly deployment-friendly.</p>