创新背景

当今自动驾驶汽车和无人驾驶飞机的图像识别技术依赖于人工智能：计算机基本上自学识别物体，如狗、过街行人或停车的汽车。问题是，运行人工智能算法的计算机目前对于手持医疗设备等未来应用来说太大和太慢。

创新过程

斯坦福大学的研究人员设计了一种新型的人工智能相机系统，可以更快、更节能地对图像进行分类，有朝一日，这种系统可以小到足以嵌入设备本身。

Wetzstein和研究生Julie Chang将两种类型的计算机结合在一起，创造了一种专为图像分析设计的光电混合计算机，朝着新技术迈出了一步。

原型相机的第一层是一种光学计算机，它不需要数字计算的高功率数学。第二层是传统的数字电子计算机。光学计算机层通过对图像数据进行物理预处理，以多种方式对其进行过滤，否则电子计算机必须在数学上进行过滤。由于滤光在光线通过定制光学器件时自然发生，因此该层以零输入功率运行。这为混合系统节省了大量的时间和能量，否则将被计算消耗。

结果是大大减少了计算，减少了对内存的调用，并大大减少了完成过程的时间。跳过这些预处理步骤后，剩下的分析将以相当大的优势进入数字计算机层。

在速度和精度方面，该原型可与现有的纯电子计算处理器相媲美，这些处理器被编程为执行相同的计算，但大大节省了计算成本。

虽然他们目前的原型布置在实验室工作台上，很难归类为小型，但研究人员表示，他们的系统有一天可以微型化，以适合手持摄像机或无人机。在模拟和真实实验中，该团队使用该系统在自然图像设置中成功识别飞机、汽车、猫、狗等。

除了缩小原型机之外，斯坦福计算成像实验室的Wetzstein、Chang和同事们正在寻找方法，使光学元件能够进行更多的预处理。最终，他们的更小更快的技术可能会取代现在帮助汽车、无人机和其他技术学会识别周围世界的后备箱大小的计算机。

创新价值

未来这种更小、更快的技术可能会取代现在汽车、无人机和其他技术的后备箱大小的计算机去学会识别周围世界。

创新关键点

原型相机的第一层是一种光学计算机，它不需要数字计算的高功率数学。第二层是传统的数字电子计算机。光学计算机层通过对图像数据进行物理预处理，以多种方式对其进行过滤，否则电子计算机必须在数学上进行过滤。由于滤光在光线通过定制光学器件时自然发生，因此该层以零输入功率运行。这为混合系统节省了大量的时间和能量，否则将被计算消耗。

New AI camera for more efficient image classification

Researchers at Stanford University have designed a new type of artificial intelligence camera system, which can classify images faster and more energy-saving. One day, this system can be small enough to be embedded in the device itself.
Wetzstein and graduate student Julie Chang combined two types of computers and created a photoelectric hybrid computer designed for image analysis, which took a step towards new technology.
The first layer of the prototype camera is an optical computer, which does not need the high-power mathematics of digital calculation. The second layer is the traditional digital electronic computer. The optical computer layer filters the image data in various ways through physical preprocessing, otherwise the electronic computer must filter it mathematically. Since filtering occurs naturally when light passes through custom optics, the layer operates at zero input power. This saves a lot of time and energy for the hybrid system, otherwise it will be consumed by calculation.
The result is that the calculation is greatly reduced, the call to memory is reduced, and the time to complete the process is greatly reduced. After skipping these preprocessing steps, the remaining analysis will enter the digital computer layer with considerable advantages.
In terms of speed and accuracy, the prototype is comparable to existing pure electronic computing processors, which are programmed to perform the same calculations, but greatly save computing costs.
Although their current prototype is placed on a laboratory bench, which is difficult to classify as small, the researchers said that their system could one day be miniaturized to suit handheld cameras or drones. In simulation and real experiments, the team successfully identified aircraft, cars, cats, dogs, etc. in natural image settings using the system.
In addition to reducing the prototype, Wetzstein, Chang and colleagues at Stanford computational imaging laboratory are looking for ways to enable more preprocessing of optical elements. Eventually, their smaller and faster technology may replace the trunk size computers that now help cars, drones and other technologies learn to recognize the world around them.