Image sensors and image signal processing (ISP) are an integral part of any current and future automotive application. Low-light conditions represent a key challenge for any such ISP solution, with noise becoming the limiting factor. For human vision applications, this noise adds speckles, blurs, and distortion to images. For machine vision applications, which depend on detection accuracy, image noise reduces the precision of object recognition. This means that for applications like reverse parking, surround view, and electrical mirrors, low light conditions are one of the biggest obstacles to safe operation.
These types of real-time edge applications require sophisticated denoising algorithms that meet performance requirements while staying within tight power and area constraints. We will describe an AI denoiser solution implemented entirely in software and executed on embedded processors and takes two orders of magnitude fewer compute cycles compared to existing approaches of similar denoising quality. Multiple benchmark examples to illustrate its performance will be presented.