To address issues such as artifacts and blurred details in super-resolution images of chip surface defects generated by traditional super-resolution generative adversarial network （SRGAN), this paper proposes a method for chip surface image super-resolution reconstruction based on an improved SRGAN. Firstly, the pyramid context extraction module (PCEM) is introduced into the generator network. This module effectively integrates multi-level information and details of the image through the pooling and concatenation of multi-scale feature maps, combined with a contextual attention mechanism. This significantly enhances the super-resolution reconstruction performance, resulting in clearer details in the generated images. Secondly, the rectangular self-calibration module (RSCM) is introduced into the generator network. By modeling context in horizontal and vertical directions, it adaptively calibrates key regions with rectangular geometric features, thereby effectively extracting the structural edges and directional textures of chip images and enhancing the model’s ability to represent chip structural information. Finally, the mobile network version 3 Small (MobileNetV3Small) is introduced into the adversarial network. This makes the super-resolution reconstruction model more lightweight, lowers hardware deployment requirements, and simultaneously strengthens feature extraction capabilities. Experimental results demonstrate that the improved algorithm increases the peak signal-to-noise ratio (PSNR) by 5.53 dB and improves the structural similarity index (SSIM) by 4.2%, effectively enhancing image quality. This provides a novel solution for the deployment of high-precision chip surface defect detection technology in industrial scenarios.