计算摄影--google相机的王者之道

Posted 2021-05-02 大话成像

这几天DXO的手机相机兵器谱实在是热闹，一个月之内，榜首三易其主。

先是iphone8 放了大招干掉了google pixel，然后没几天，三星note8 又横空出世来了个并列第一。

苹果三星这两个硬件巨人挤在铁王座上还没有把座位捂热乎，软件巨头google的pixel 2又一下夺回了手机相机的王者荣耀。

Dxo的评价是：google pixel 2 创下了手机相机的新记录

Test summary

The Google Pixel 2 not only takes the DxOMark Mobile scoring crown with a stunning Overall score of 98, but its video performance improves by 5 points over the original Pixel, putting it at 96 and at the top of our video ranking. The Pixel 2’s photo performance is really impressive at 99, and just one point short of the Samsung Galaxy Note 8, which scored an amazing 100 in our photo category.

google pixel 2的软件ISP方案，居然在视频质量上提升了5分，达到了96分，登上视频质量的王座，这真是太强大了！！！不得不为google的camera 团队点赞！！！

这是google 相机的基本特性：

• 1/2.6-inch     12MP sensor

• f/1.8 aperture

• Dual-Pixel     Autofocus

• Optical Image     Stabilization

• HDR+     technology

• 软件方案的人像模式

 

表面看起来没有什么特别，那么google是如何打败竞争对手的呢？

在产品发布会上，google的回答是：

It's a combination of the hardware with its computational intelligence.

 

Google 的camera研发团队的负责人Marc Levoy， 以前是斯坦福大学计算机图形实验室的教授，在接受记者采访时阐述了 google 的相机技术方向----

"I think the excitement is actually just starting in this area, as we move away from single-shot hardware-dominated photography to this new area of software-defined computational photography."

 

computational photography，计算摄影 , 就是google 相机的王者之道。

 

那么计算摄影到底能做什么呢？优势在哪呢？

 

计算摄影是基于传统数字成像而衍生的新的成像技术：包括四个主要部分：

 

 

Computational Processing，

Computational Imaging/Optics ,

Computational Sensor ,

Computational Illumination.

 

Computational Processing包含以下子技术：

Mosaicing,

Matting,

Super-Resolution,

Multi-Exposure HDR,

Light Field from

Mutiple View,

Structure from Motion,

Shape from X.

 

Computational Imaging/Optics 包含以下子技术：

Coded Aperture,

Optical Tomography,

Diaphanography,

SA Microscopy,

Integral Imaging,

Assorted Pixels,

Catadioptric Imaging,

Holographic Imaging.

 

Computational Sensor  包含以下子技术：

Artificial Retina,

Retinex Sensors,

Adaptive Dynamic Range Sensors,

Edge Detect Chips,

Focus of Expansion Chips,

Motion Sensors.

 

Computational Illumination包含以下子技术：

Flash/no flash,

Lighting domes,

Multi-flash for depth edges,

Dual Photos,

Polynomial texture Maps,

4D light source

 

具体到google pixel 2 相机，我们能看到哪些计算摄影的应用呢？

 

1： 快速对焦。

google可以利用digital refocusing 技术实现超快速对焦。

同样是使用了dual pixel PDAF的硬件，google的对焦做的比别人更快更稳定，在DXO评测中拿到了96分的最高分。

2. 光学\数字防抖+去模糊算法

3 .更好的HDR+ 技术

 

The traditional way to produce an HDR image is to bracket: you take the same image multiple times while exposing different parts of the scene, which lets you merge the shots together to create a final photograph where nothing is too blown-out or noisy. Google's method is very different — HDR+ also takes multiple images at once, but they're all underexposed. This preserves highlights, but what about the noise in the shadows? Just leave it to math.

 

google的HDR方法是欠曝光HDR多帧融合，可以更大的避免抖动，可以降低图像配准的难度，然后利用强大的去躁算法去除低曝光带来额外的噪声。

 

4. 更好的低光照拍照性能

google also claims that, counterintuitively, underexposing each HDR shot actually frees the camera up to produce better low-light results. "Because we can denoise very well by taking multiple images and aligning them, we can afford to keep the colors saturated in low light," says Levoy. "Most other manufacturers don't trust their colors in low light, and so they desaturate, and you'll see that very clearly on a lot of phones — the colors will be muted in low light, and our colors will not be as muted." But the aim isn't to get rid of noise entirely at the expense of detail; Levoy says "we like preserving texture, and we're willing to accept a little bit of noise in order to preserve texture."

 

一般消费相机在低光依然希望能保持色彩饱和度，但是就容易出现明细的彩色噪声，google的优秀降噪算法可以避免这种窘境。

 

5. 单摄像头软件方案的人像模式拍照，这个功能有点打脸双摄像头及深度传感器，google用一个摄像头就实现背景虚化这种人像模式的拍摄，而不需要额外的硬件。

 

关于计算摄影，还有很多在成像科学上的应用，比如光场相机，编码快门，反射分割，结构照明等等。

总的来说，google的软件相机方案具有更加灵活，更低成本，更快实现产品迭代的优点，在目前的相机竞争中可以说独树一帜，而且这种软件方案，让安卓相机不必完全依赖高通的平台更新节奏，能够最大的实现产品快速优化，值得手机相机开发者考虑采用。