2022年4月6日
By 塞吉奥Grce 首席执行官iSize
聚光灯

把不好看的变成好看的

塞吉奥Grce discusses BitClear, an artificial intelligence based denoising technology.

Consider the life-cycle of a piece of “viral” media. 这可能是对新闻故事的罕见洞察, 意想不到的艺术成就, 或者只是一个有趣的动物. But it has spread around the media-sharing social sites.

考虑到现在相机的质量, 甚至在手机里, it may well have started with quite pristine video imagery. But each time it gets shared it gets re-encoded. At the point when it has the potential for viral success – and therefore valuable monetisation – the quality has often degraded significantly. We are all well aware of concatenated encoding (a.k.a. transcoding), like blurring, blocking and ringing.

For many, this poor quality makes it unwatchable, however interesting the content. It is a paradox: the more successful a video clip on social media, the fewer people will tolerate distortion and artifacts to watch it.

What is needed is a process that rolls back the effects of all those encodings and recovers quality as close as possible to that of the original video. 不幸的是, conventional wisdom is that video quality lost after each transcoding can never be recovered.

iSIZE specialises in the application of deep learning in video delivery, 所以我们调查了这个问题. 结果是BitClear, an artificial intelligence based denoising technology that is specifically tuned to remove the effects of compression.

BitClear is a bespoke neural architecture that is trained to disentangle the noise from the data manifolds, i.e., 去除压缩噪声和伪影, and retain or recover as much of the original content as possible. Unlike generic denoising technology that targets film grain or interlacing artifacts, BitClear targets the compression artifacts of typical MPEG or AOMedia encoders.

But while it learns the noise signature of the various common encoding standards, BitClear does not need to know the history of the specific asset. It can process any highly-compressed content and produce a much higher quality version at twice the resolution of the original input. That, in turn, improves the value of the asset. 在英特尔至强处理器上, the fastest BitClear models achieve processing speeds of 30 frames a second when producing content at 1080p resolution. That speed makes it applicable for a wide range of content in both live and VoD processing pipelines.

Its bespoke neural network architecture also means that it is readily scalable, up to very high volumes of parallel processing. Its architecture means it can be implemented in the cloud or on premise, 并且可以在cpu上运行, GPUs or custom hardware that supports neural network inference.

Taken together, this makes BitClear ideal for its primary target market, e.g. 社交媒体网络中的短格式内容, or when seeking to maximise the potential value from popular user-generated content or legacy-encoded content at lower resolution. At present, the solution is in evaluation trials with a number of major UGC distribution platforms.

The underlying AI principles of BitClear though, can potentially be applied to other applications which require the separation of image from noise and distortion. That could include cleaning archive content, for example. It would also be applicable where delivery bandwidth is limited by infrastructure, for example in broadcast contribution feeds over mobile streaming, 或者是游戏, 现场直播的体育, 娱乐和虚拟现实. IoT applications like security start with very high image quality, 1080p或更高, so it makes sense to preserve that content whatever the delivery bandwidth, 尤其是通过移动网络发送时. There is broad scope for realtime or near-realtime image enhancement and de-noising across many industries.

There are many reasons why media businesses would want to reduce or eliminate the effects of multi-generation encoding and other distortions. 可能是内容有新闻, 历史或文化意义, or it could simply be the business view that the higher the quality, 收入越大.

总之, BitClear achieves excellent results in removing artifacts without impacting the creative intent of the original content. It does it without user intervention or subjective decision-making: it is suitable for an entirely automated workflow. And it does it at speed and at scale, on readily available hardware, on-premise or in the cloud.

http://www.isize.co/

这篇文章是赞助内容