Image super-resolution reconstruction algorithm based on multiscale residual dense network and improved transformer

Image super-resolution reconstruction seeks to restore high-resolution details from low-resolution images. A two-branch parallel algorithm combining a Multiscale Residual Dense Network (MRDN) and an improved Transformer is suggested. First, the channel attention mechanism is fused with MRDN, using parallel convolutional kernels to extract texture and structural features in different receptive fields, enhancing the interlayer information flow through dense connections, and introducing channel attention to adaptively recalibrate the feature channel weights. Based on this, a dual-branch alignment module grounded in deformable convolution is constructed, and a Hybrid Attention Module (HAM) is suggested to fuse local and global features. Experimental outcomes reveal that on three datasets, the peak signal-to-noise ratios of the suggested algorithm under the ×4 magnification factor reach 32.19, 31.67, and 30.51 dB. The outcomes reveal that the proposed algorithm can maintain high fidelity, offering an efficient and feasible technical solution for image super-resolution tasks.