U-Net-based image restoration for hybrid refractive–diffractive optical structures

Hybrid refractive–diffractive imaging systems can effectively suppress chromatic aberration, but the captured images still suffer from residual blur, low contrast, and photon loss, which limit the practical image quality. In this work, we propose a deep-learning-based image restoration framework for hybrid optical imaging. A U-Net-based reconstruction network is employed to learn the inverse mapping from degraded hybrid measurements to high-quality images. By combining multiscale feature extraction, skip connections, and self-attention, the network suppresses chromatic artifacts and diffraction-induced blur while preserving fine structural details. Experimental results on the DLMD dataset show that the proposed method achieves an average PSNR of 34.7 dB, an SSIM of 0.912, and an inference time of 0.064 s per image, outperforming existing reconstruction models in both reconstruction quality and computational efficiency. The proposed approach provides an efficient and robust solution for enhancing image quality in hybrid refractive–diffractive imaging systems.