Artificial Intelligence (AI) is disrupting the way business is done, from everything from retail to manufacturing and automotive to banking. Enterprises and their R&D teams may be developing different products and services, but all are racing on one common ground to combine AI with various application scenarios to meet evolving market demands.
AI is empowered by data. Machines utilize intelligent algorithms to analyze, extract, and compare the acquired data and, subsequently, perform a series of corresponding automated processing flow, based on analysis results. Of all the different types of data, image data is the most objective and direct expression of the real world. Therefore, to fully activate the power of AI in any application – such as high-end industrial automation, common code readers, drones, or cleaning robots – the ability to capture comprehensive and accurate image data is the most critical of all.
And as one of the most important means to acquire image data, CMOS image sensors (CIS) are slated to achieve faster development and reach a broader market in the AI era. Currently, there are two mainstream technologies applicable to CMOS image sensors: FSI and BSI.
FSI (front side illuminated) is designed in accordance with the semiconductor manufacturing process. For pixels, light enters between the front metal wiring and then focuses on the photosensitive area (photodiode). The performance of the FSI can meet the requirements for larger pixels because the ratio between the height of the pixel's optical stack and the pixel area is small, and the photosensitive area is relatively guaranteed. However, as a sensor's resolution gets higher, the pixel size is reduced, the fill factor gets smaller, the optical path is elongated, and the metal wiring reflection absorption loss is larger, which limits the performance of the sensor.
BSI (back side illuminated), on the other hand, separates electrical components from light, allowing the optical path to be independently optimized, avoiding absorption, reflection and flare of the FSI metal wiring layer. In addition, the optical stack in the BSI pixel is greatly reduced. When compared with FSI, BSI has a large fill factor of almost 100 percent. Therefore, BSI can achieve higher QE (quantum efficiency).
The manufacturing process of FSI is simple, low-cost, high-yield and requires less technique compared to the complexities and difficulties of BSI technique. In the early development of BSI, yield was a big obstacle that needed to be overcome. But thanks to progress in the development of semiconductor technology, BSI technique is becoming more and more mature. Yield also rises fast.
The CMOS image sensor using BSI technology is suitable for applications that need high resolution with limited optical and pixel size, high sensitivity and strong low-light performance. For example, high-end security and surveillance, factory automation and mobile phones. Looking ahead, BSI is the trend of CIS development.
As a separate trend, global shutter is fast-growing, in terms of adoption in comparison to rolling shutter. Achieving excellent real-time performance without the jelly effect is contributing to the popularity of global shutter CIS, especially in the fields of AI and machine vision applications. As process technology becomes more mature and cost continuously comes down, it is certainly within reason to expect the market demand for global shutter CIS will increase swiftly.
For example, factory automation is one field in high demand of AI, where high-performance, industrial-grade global shutter CIS can be widely applied to improve inspection, quality control, optical character recognition, robot arm guidance and other automated functions. In drone applications, obstacle avoidance and optical flow are major focuses for CIS.
But CIS technology can be used more broadly beyond merely industrial applications. From barcode scanning at a grocery store to a lane departure warning from an automobile, CIS technology is greatly and increasingly improving quality of life.
It is a recent technological breakthrough to combine a BSI pixel design with global shutter technology, which in result provides superior signal-to-noise ratio, higher sensitivity and greater dynamic range. Such technology can be applied to a wide range of commercial products, including smart barcode readers, drones, smart modules (gesture recognition/vSLAM/depth information/optical flow) and other image recognition-based AI applications, including facial recognition and gesture control.
Compared with conventional CMOS image sensors using Multiple-Exposure HDR technology, single-frame HDR Global Shutter technology is more suitable for image recognition-based AI applications. When combined with DVP/MIPI/LVDS interface, this technology can be adapted to various types of SoC platform.
In the age of the fourth industrial revolution, more and more customers are looking to deploy cutting-edge AI algorithms and applications on embedded platforms to speed up system response and make devices more intelligent. Accurate image data can help AI/deep learning algorithms analyze and see the world more clearly.
About the author
Leo Bai is the General Manager of Artificial Intelligence BU at SmartSens . He has earned a MS degree from Harbin Institute of Technology, China and has more than 10 years of high-end embedded intelligent imaging system design and product management experience. Leo has successfully launched three series of smart cameras that were widely used in ITS and FA field applications and is a leader in the development trend of intelligent applications.