Deep dive into New Huawei AI ISP

Huawei AI ISP Introduction

Huawei AI ISP Introduction: Huawei Yueying AI ISP

Huawei HiSilicon Kirin chip is hampered, but in other areas, Huawei HiSilicon did not stop there and recently launched the new Huawei AI ISP, a new generation of intelligent image processing engine for IoT smart terminals, which breaks through the limits of traditional ISP image processing and pushes image quality to new heights.

Huawei AI ISP Introduction

ISP is the first link in the image processing of the raw data input from the image sensor, covering dozens of image signal processing algorithms. To adapt to different lenses and sensors, hundreds of parameters need to be optimally adjusted to ensure that the ISP is both versatile and flexible.

At the same time, the algorithms of many modules in the ISP interact with each other, and the collaboration of algorithms requires a lot of debugging work, which is extremely complex and challenging.

In the face of increasing scene complexity, traditional image processing techniques based on filtering are becoming increasingly difficult to achieve significant improvements in image quality in dark light, wide dynamic scenes, etc. Moreover, as the parameter library gradually becomes larger, debugging becomes more and more difficult, the development cycle gradually lengthens, and the ROI of image quality improvement becomes lower and lower.

In recent years, the use of AI for image enhancement has gradually become a new industry research hotspot and has made significant progress, especially the use of AI for the optimization of still photos, which has been widely used in high-end cell phone cameras photography and has obtained good results.

Taking advantage of its accumulation and experience in traditional ISP image quality technology, as well as its advantage in end-side AI architecture and optimization, Shanghai HiSilicon has created a new generation of intelligent image processing engine – Yueying AI ISP, which deeply integrates traditional image quality optimization experience with AI technology, achieving a leap in five aspects compared to HiSilicon’s previous-generation flagship ISP.

Ultra-sensitive noise reduction: Intelligent noise reduction based on deep learning, four times higher signal-to-noise ratio

Traditional NR noise reduction technology uses multi-level time domain/space domain filtering, and the filter design incorporates a variety of heterogeneous types, the gains have gradually decreased.

HiSilicon uses neural networks to deeply learn the distribution characteristics of noise and signal in a large number of low-light scenes pictures and train a set of intelligent noise reduction algorithm models, and Huawei AI ISP can intelligently distinguish signal and noise in images to achieve intelligent noise reduction in low-light scenes.

According to the test, compared to the previous generation of flagship ISPs, the signal-to-noise ratio of images in low-light scenes can be improved by 4 times.

Multi-spectral fusion: Using dedicated hardware acceleration to achieve 0.01Lux low-light full color

In low-light scenes, the RGB sensor’s light-sensitive capability drops dramatically, resulting in a rapid decline in color image clarity and an inability to take into account both image detail and color.

In the night traffic management scenario, it is often necessary to install flashing lights at the gates to supplement the lighting, which can easily cause light pollution and even affect the safety of drivers.

Multispectral fusion technology by fusing visible and infrared light perception information, without increasing light pollution at the same time, to solve the problem of loss of detail caused by the lack of visible light sensitivity in a low-light environment, to achieve the best balance of image color and detail, and finally make the image effect to 0.01 Lux dark light full color, rich in detail texture.

Super Wide Dynamic Range: Self-adapting to complex changes in light, 12dB increase in dynamic range

Wide dynamic technology is an important technology to improve image quality in backlight scenes. For example, when driving in and out of the tunnel entrance, due to the huge contrast between the internal and external light intensity, how to quickly track the changing light and accurately adaptive exposure of the target is a very important technology for automatic and assisted driving safety.

Super Wide Dynamic technology is based on intelligent recognition of the target scene and adaptive multiple exposures of the value area. Through intelligent algorithms fusing image data of long, medium, and short frames with different exposure times, the image obtains a perfect blend of dark and light areas: bright areas are not overexposed, dark details are not lost, backlit scenes are seen clearly, and dark areas are full of color and detail.

According to the test, compared to the previous generation of flagship ISP, Huawei AI ISP can improve the dynamic range in backlight scenes by 12dB, and track the scene more quickly to effectively avoid instantaneous over darkness or overexposure of the image, and converge to a stable state within 3 frames.

Super Steady Image: 6-axis digital stabilization, video imaging stability greatly improved

Consumer cameras are widely used in the shooting of sports scenes, but the vibration of the camera in motion can affect the final imaging effect. Visual perception devices are installed in autonomous or assisted driving vehicles, and such shaking causes blurred images that may even affect driving safety.

Based on IMU 6-axis sensing information to intelligently analyze motion posture and correct dynamic video through precise compensation, Yueying Super Steady Image Technology effectively reduces the impact on video imaging due to external environmental shaking or jitter, and can significantly improve video stability in various vibration environments.

Multi-dimensional perception: Multi-sensory cross-border integration, stimulating new potential of intelligent vision

Yueying supports the fusion of RGB and radar, infrared thermal imaging, 3D structured light, and other multi-dimensional perception capabilities, matching the needs of infrared temperature measurement, intelligent payment, sweeping robot, radar video integrated machine, and other products, which can be widely used in traffic management, financial payments, industrial vision, and other business scenarios.

Vision and AI are key technologies in many areas of new infrastructure and the most important fundamental capabilities of intelligent terminals. Based on the vision industry, Shanghai Haise has deep experience and industry-leading core capabilities in end-side vision perception and AI computing, insisting on investment and continuous innovation to provide industry partners with the most leading technologies and solutions. With the deep integration of vision and AI capabilities, the Hesi VC-IF AI ISP will provide more superior vision capabilities for intelligent terminals in thousands of industries, creating more value for society and people in many aspects such as urban management, industrial production, smart life, and driving safety.

HiSilicon.

Source 1, Source 2

Exit mobile version