GAN-Based Image Restoration for Enhancing Object Detection in Projector-Camera Systems

Abstract

Projector-camera systems are widely utilized in fields such as augmented reality (AR), education, and healthcare, offering intuitive interaction by projecting digital content onto physical surfaces and detecting objects in real time. However, light emitted from the projector can cause severe color distortions, degrading the performance of color-based object detection. In this study, we propose a Generative Adversarial Network (GAN)-based image restoration model designed to correct such projection-induced distortions. The model incorporates a color condition vector derived from the projector's illumination, attention and residual blocks in the generator, and a similarity map module in the discriminator, optimized with WGAN-GP and perceptual losses. By restoring distorted images to their original appearance, the proposed method improves detection accuracy without retraining existing object detection models. Experimental results on a 50,000-image projector-camera dataset demonstrate that our approach outperforms representative restoration models-including Autoencoder, SRCNN, U-Net, ResNet50, and DnCNN-across all quantitative metrics (e.g., LPIPS 0.078, CIEDE2000 5.766, SSIM 0.903). Furthermore, object detection accuracy reached 97.2% for template matching and 99.2% for YOLO, nearly matching the performance on original images. These results confirm the effectiveness of the proposed method in enhancing object recognition under projection-based environments and indicate its potential for robust deployment in diverse AR and SAR applications.