DifuzCam replacing camera lens with a mask and a diffusion model for generative AI based flat camera design

DifuzCam 用掩模和扩散模型取代相机镜头,用于基于生成式人工智能的平面相机设计

阅读:1

Abstract

Recent advances in lensless, flat camera designs hold the promise of significantly reducing size and weight by replacing bulky lenses with thin optical elements that modulate incoming light. However, recovering high-quality images from the raw sensor measurements of such systems remains challenging. We address this limitation by introducing a novel reconstruction framework that leverages a pre-trained diffusion model, guided by a control network and a learnable separable transformation. This approach delivers high-fidelity images, achieving state-of-the-art performance in both objective and perceptual metrics. Our method achieves 20.43 PSNR, 0.612 SSIM, and 0.237 LPIPS on the FlatNet dataset, representing improvements of 9.6%, 18.1%, and 26.4% respectively over the previous state-of-the-art FlatNet method. Additionally, the text-conditioned nature of the diffusion model enables optional enhancement through scene descriptions, particularly valuable for compact imaging systems where user input can help resolve reconstruction ambiguities. We demonstrate the effectiveness of our method on a 8 flat camera, paving the way for advanced lensless imaging solutions and offering a robust framework for improved reconstructions that is relevant to a broad range of computational imaging systems.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。