핵심 개념
Global Search Optics (GSO) is a comprehensive end-to-end lens design framework that automatically explores the solution space to design compact computational imaging systems with superior imaging quality compared to traditional methods.
초록
The paper introduces Global Search Optics (GSO), a novel framework for the design of compact computational imaging systems. GSO consists of two key components:
-
Fused Optimization Method for Automatic Optical Design (OptiFusion):
- OptiFusion combines Simulated Annealing (SA), Genetic Algorithm (GA), and ADAM to automatically search for diverse initial optical system structures based on traditional optical design metrics.
- This addresses the limitation of traditional methods that often fail to identify initial lens designs closely approximating the global optimum.
-
Efficient Physic-aware Joint Optimization (EPJO):
- EPJO features an advanced differentiable simulation model that incorporates differentiable ray tracing, patch-wise convolution, and an Image Signal Processing (ISP) pipeline.
- EPJO also employs customized memory-efficient techniques to enable parallel joint optimization of the initial structures discovered by OptiFusion and image reconstruction models.
- This allows for the selection of the jointly optimal solutions based on the final reconstructed image quality metrics, while considering the intricate physical constraints of optical systems and the categorical nature of glass materials.
The paper validates GSO's superior global search capability through comparison with traditional methods of initial structure determination. It also demonstrates the weak correlation between traditional optical design metrics and final reconstructed image quality metrics. Definitive experiments are conducted to establish the joint design approach's substantial enhancement in the performance of computational imaging systems over the separate design mode.
통계
The paper reports the following key metrics:
For the 3P-I design specification, GSO improves PSNR by 0.25dB~0.38dB, SSIM by 0.0072~0.0103, and LPIPS by 0.0026~0.0053 compared to the CODE V assisted joint design method.
For the 3P-II design specification, GSO enhances PSNR by 0.25dB~0.40dB, SSIM by 0.0089~0.0127, and LPIPS by 0.0033~0.0067 compared to the CODE V assisted joint design method.
인용구
"Global Search Optics (GSO) is a comprehensive end-to-end lens design framework that thoroughly and autonomously explores the solution space for compact computational imaging systems."
"The paper validates GSO's superior global search capability through comparison with traditional methods of initial structure determination, and demonstrates the weak correlation between traditional optical design metrics and final reconstructed image quality metrics."
"Definitive experiments are conducted to establish the joint design approach's substantial enhancement in the performance of computational imaging systems over the separate design mode."