Title: Independent Positioning Of Auxiliary Information In Neural Network Based Picture Processing
Application Date: 08/11/23
Grant Date: 24/01/2025
Patent Number: 558691
Patent Holder: Huawei Technologies Co., Ltd
Nationality: Foreign
Field of Invention: Communication
What is this patent about?
This patent concerns the use of neural networks to compress and decompress pictures and video.
It describes a method for moving and combining information between different stages of a neural network during video processing. In particular, information obtained from a later stage of the network is used together with other information to generate input for an earlier stage.
The patent covers both decoding and encoding. In decoding, data from a compressed video is processed through a neural network to produce picture data. In encoding, a picture is processed through a neural network and the resulting feature data is included in a video.
The patent primarily concerns algorithmic techniques for picture and video compression and processing.
What does the patent claim?
Claim 1 describes a method for processing feature data from one or more pictures using a neural network with at least two stages. First, the method obtains data from a bitstream. It then obtains other data from a later stage of the neural network, where that data is based on processing that has already taken place. The two sets of data are then used together to generate input for an earlier stage of the neural network.
The patent then claims more specific versions of this basic arrangement. For example, Claim 3 describes the first data as a “prediction error” and the second data as a “prediction”. In video compression, a prediction is an estimate of what a picture or part of a picture should look like, while the prediction error is the difference between that estimate and the actual picture data. The patent also claims the use of motion or spatial information to generate the prediction. This can include motion vectors, which indicate how parts of a picture have moved between frames. The claims specify how the prediction and prediction error are combined. One claimed method uses element-wise addition, where corresponding values in the two sets of data are added together.
Another part of the patent deals with probability-model data from one stage of the neural network to be used in entropy decoding. In other words, information generated at one point in the network is used to help decide how compressed data should be decoded. The same basic approach is claimed for encoding. Claim 14 covers processing a picture through a multi-stage neural network, combining information from different stages, using that information to generate input for an earlier stage, and then including the resulting feature data in the bitstream.
The patent also claims apparatuses with processing circuitry configured to carry out these operations. More importantly, it expressly claims a computer program stored on a computer-readable, non-transitory medium, where running the program causes a processor to perform the claimed methods.
Taken together, the patent claims a particular way of processing video using a neural network: information from different stages of the network is brought together, combined in specified ways, and then used to affect processing at another stage.
Three-Part Test under 2016 CRI Guidelines
The 2016 Computer Related Inventions (CRI) Guidelines require patent examiners to follow a three-part test to determine whether a computer-related invention is patentable.
1) The claim to be examined in substance
The actual contribution of this invention is a neural network method that combines feature data from different stages of processing to improve video compression efficiency. In substance, the invention is about how software processes data within a neural network. This is an algorithm-based invention since the crux of the patent entirely revolves around the neural network model.
2) Denied unless the actual contribution extends beyond a computer programme or algorithm
The invention is presented as a technical process used in video coding rather than merely as a mathematical method, business method, or abstract algorithm. Because it is applied in the context of video compression, it could be said that the actual contribution relates to a computer programme alone.
3) Denied unless the claim involved computer programmes combined with novel hardware
The patent relies on a neural network implemented through software. It does not disclose any new or specialised hardware. Instead, it uses conventional processing circuitry to carry out the claimed method. The inventive contribution lies entirely in the computer programme, rather than in any novel hardware combined with the software.
Since the invention is essentially software running on standard computing hardware, it would not have been considered patentable under the 2016 CRI Guidelines.
Why is this an egregious software patent?
The claimed innovation is a software-based neural network technique for improving video compression. The invention itself does not introduce any new hardware. The patent does not claim a new processor, memory architecture, camera, video chip, or other specialised hardware. The improvement comes entirely from the way the software processes information, as the patent primarily concerns an algorithm.
Under the 2016 CRI Guidelines, this would not have qualified for a patent because the contribution lies solely in a computer programme implemented on conventional hardware. However, under the 2025 CRI Guidelines, such software-based methods are granted patents in a manner that goes against the legislative intent of Section 3(k) of the Patents Act, 1970.
Under the 2025 CRI Guidelines, greater emphasis is placed on technical effect and technical contribution when assessing computer-related inventions. A software-based technique used for video compression can therefore be considered in terms of the technical result it produces.
The problem with the technical effect test, as identified in our previous blogs on the subject (Part I, Part II) is that the production of a technical result does not necessarily change the nature of the underlying software. In this patent, the claimed method still concerns the organisation and processing of information within a neural network. The patent does not claim a new machine for video compression. It claims a particular way of processing information: obtaining data from different stages of a neural network, combining that data, and using it at another stage. It also claims the computer program that carries out these operations. Therefore, the patent in question falls squarely within the bounds of “software patent.”
Finally, the detailed reasoning of the Patent Controller is not available on the Indian Patent Office website, as the patent was granted without a hearing. This makes it difficult to assess how the Patent Office considered the Section 3(k) objection and the substance of the claimed invention.
