Efficient Image Super-Resolution Guided by Object Detection

Abstract

To address the challenges of reduced target recognition accuracy and insufficient real-time performance caused by low image resolution in traditional tower assembly construction, this paper proposes an efficient image super-resolution method guided by object detection. The proposed approach employs YOLOv8 to rapidly locate target regions and introduces a discriminative image segmentation module to divide the image into key and background regions. These regions are then processed separately using a specially designed Super-Resolution Guided Efficient Detail Recovery Network (SGERN) and a Background Lightweight Super-Resolution Network (BLSRN), thereby enhancing image quality while significantly reducing computational cost. The SGERN integrates Generative Adversarial Networks (GAN), a simple gating mechanism, and re-parameterization modules to achieve high-quality reconstruction of fine textures in key regions. Experiments conducted on a self-built dataset of hoisting components and public datasets (UDM10 and Urban100) demonstrate that the proposed method outperforms mainstream approaches in terms of PSNR and SSIM. Moreover, it achieves good real-time performance while ensuring high visual quality in the target areas. This study provides a novel and practical solution for target recognition and image enhancement in complex construction scenarios, with promising engineering application value.