Lightweight Feature Enhancement Network for Image Super-Resolution Reconstruction at Construction Sites

Constrained by environmental factors and equipment limitations, the images obtained from construction sites suffer from low resolution and strong noise interference. This lack of detail severely influences the accuracy and effectiveness of construction site inspections. To respond to these challenges, a lightweight feature enhancement network (LFEN) is proposed for image super-resolution reconstruction at construction site. Firstly, a dilated cross-convolution block is put forward to enhance image noise immunity and reconstruction effectiveness at edge locations. Secondly, the traditional non-local attention mechanism is improved to reduce the number of parameters while focusing on global visual correlations. Finally, a multi-path dense residual cascade method is used to build the entire reconstruction network. Furthermore, a noise immunity loss function is employed to enhance overall network performance. Extensive experiments are carried out using real-world construction site datasets. The network achieves a PSNR value of 30.42 for 4 \(\times \) $\times$ super-resolution on the test set, with a parameter number of 6.22M. This signifies the reconstruction quality with reduced parameters. The experimental results demonstrate that the proposed network substantially enhances visual perception in image super-resolution reconstruction. The proposed network surpasses current mainstream algorithms at construction site, and has wider applicability.