Laser-induced glare effect and image reconstruction via a frequency-aware transformer

Laser provides a low-cost and potent threat to optoelectronic imaging systems, capable of saturating sensor pixels and interfering with visual information. Reconstructing laser-glared images by algorithm presents challenges due to the complex laser patterns that mask scene information. Herein, we experimentally investigate the laser-induced glare effect in the electro-optical imaging process with diverse jamming illumination configurations. The results show that laser glare creates a hybrid image degeneration characterized by complete occlusion surrounded by a halo, which in turn causes both information loss and color shifts. High-spatial-frequency speckle noise scattered within an optical lens is observed in glare due to its high intensity and strong coherence, distinct from the glare scenario of natural light. To restore the laser-glared images, we propose a frequency-aware transformer network, referred to as LaserFormer. LaserFormer implements a dual-path strategy that concurrently processes spatial and frequency information, enabling interaction learning at both the pixel and patch levels. A sparse feature extractor is integrated into the network to capture critical high-spatial-frequency components and structural details, thereby improving both the visual quality and contextual accuracy of the reconstructed images. Moreover, we establish, to our knowledge, the first real-world dataset of laser-glared images, including diverse laser-jamming configurations. Experimental results demonstrate that LaserFormer outperforms existing baseline models based on convolutional neural networks and transformers in restoring laser-glared images. Specifically, LaserFormer achieves enhancements of 1.43% in PSNR and 0.83% in SSIM, along with significant reductions of 13.37% in FID and 14.37% in LPIPS, indicating its superior performance in both distortion mitigation and perceptual quality restoration. Our work offers a deployable and efficient solution for enhancing the robustness of optoelectronic imaging systems operating in complex lighting environments, such as autopilot, security monitoring, and reconnaissance applications.