Image restoration methods for simple optical systems based on deep learning

Traditional optical system design applies the modulation transfer function (MTF) and spot diagram as optimization indicators, often leading to complex lens group structures. Although simple optical systems offer significant advantages of small size and light weight, their severe aberration problem causes a sharp degradation in imaging quality. Deep learning can compensate aberration problem without increasing system structural complexity. To meet the requirements of lightweight design and high-quality imaging, this study constructs a MIMO-UNet restoration framework enhanced with coordinate attention and deformable convolution. Experiments show that in a 100 mm doublet optical system, the improved model achieves an average PSNR of 35.42 dB. A comparative experiment with the Double Gauss system confirms that the proposed method achieves significant improvement in imaging quality with fewer lenses, and the imaging result is almost the same as that of the Double Gauss system.