End-to-end single-lens imaging system design via differentiable ray tracing

The minimalist optical imaging system has shown extensive potential in machine vision and low-altitude remote sensing due to its advantages of lightweight design and cost-effectiveness. However, the existing single-lens computational imaging methods struggle to attain the global optimal solution of the system, due to the separate optimization of the frontend optical design and the back-end computational reconstruction, thereby limiting the further improvement of image reconstruction performance. In this paper, we propose an end-to-end joint automatic optimization single-lens imaging system based on differentiable ray tracing. A differentiable ray tracing model is established to realize the mapping between the physical parameters of the lens and the Point Spread Function (PSF), and a multi-scale fusion network is used to achieve achromatic image reconstruction of the single lens. The proposed method can achieve an image Peak Signal-to-Noise Ratio (PSNR) of 36.08dB, which is 14.69dB higher than that of the original chromatic aberration input image. We fabricated an aspheric lens and conducted prototype experiments for verification. The experimental results are basically consistent with the simulation design. The proposed method provides a novel solution for low-cost intelligent and compact optical design.