From Faces to Football Fields: How Applying AI Models to Different Application Fields Can Prove Inventive in China

As AI models become versatile and adaptable across multiple contexts and industries, questions about inventive step sit at the heart of patent examination in China. The 2023 Patent Examination Guidelines (hereinafter referred to as 'the Guidelines') included several sessions dedicated to explaining how inventive step should be examined for AI-related inventions, featuring examples related to AI algorithms, big data, and user experience¹.

In the re-examination of Beijing ByteDance Network Technology Co. Ltd.’s application entitled “Method and apparatus for processing an image” (Application No. 201810734681.2)², which was recognized as one of the Top 10 Re-examination Cases of 2025 by the CNIPA, the CNIPA focused on the following issue: when an existing algorithm or AI model is applied to a different application field compared to the prior art, how should its inventive step be judged?

Summary of the Invention

The invention relates to methods of processing captured images using a key point detection model. Its core lies in inputting a captured image of a target object—such as a sports field or other pre-defined object—into a pre-trained key point detection model that corresponds specifically to that target object. A "key point" is a pre-determined position on the target object (for instance, the corner point of a football goalpost).

The model outputs a position information set, which consists of: (1) Position coordinates indicating where the key points appear within the captured image; and (2) Visibility information reflecting the probability that the key point is actually visible in the captured image. These output data enable the system to recognize and align the same target object across multiple images.

One exemplary embodiment highlighted in the application was in the context of a football field. Using the pre-trained model of a specific football field, the method could analyze multiple frames of the same football field, determine geometric transformations across frames, and thus enable game data analysis.

Prosecution History

The claims were initially rejected for lack of inventive step over the combination of D1, D2 and common knowledge in the art under Article 22.3 of the Patent Law.

• D1 disclosed algorithms designed for detecting whether facial images were suitable for recognition, using key points to flag images where visibility was obscured. The output (probability of occlusion) was then used to test if the image quality met a threshold, functioning as a pre-screening step before facial recognition.
• D2 disclosed very similar concepts.

Amendments

During re-examination, the applicant amended the independent claims to:

• explicitly define that the target object includes a sports field; and
• clarify that the key point detection model was trained on captured images and corresponding position data of that specific object.

PRD’s Decision

Upon reviewing the amended claims and the applicant’s arguments, the board at the Patent Reexamination and Invalidation Department (PRD) (hereinafter referred to as “the Board”) revoked the examiner’s rejections, concluding that the amended claims did in fact possess an inventive step based on the reasons below.

Differences from Prior Art

The Board identified the following distinguishing features of the amended claims compared to the prior art:

• Training Data and Model Specificity: The claimed method required specific models trained for each individual object (i.e. a particular sport field), unlike D1 and D2’s models which were based on generic datasets of face images of different individuals.
• Input and Output: Inputs were multiple captures of the same object under different conditions rather than a random single facial image in D1 and D2; outputs included position coordinates and visibility information used for geometric correlation rather than quality screening
• Technical Effect: Enabled cross-frame analysis to determine the geometric transformation relationship, thereby supporting game data analytics; in contrast, D1 and D2’s purpose was to filter out low-quality images to reduce computational load in facial recognition systems.

Technical Problem Solved

Based on the distinguishing features above, the Board identified the technical problem addressed by the invention as how to determine the correspondence of the same target object across multiple captured images. This was materially different from what D1 and D2 achieved, which only addressed the quality of a single facial image.

The Board concluded that there was no technical inspiration in D1 or D2 to adapt their teachings to address the technical problem solved by the present invention. Accordingly, the amended claims were considered inventive.

EIP Thoughts

This case demonstrates both the challenges and opportunities for applicants seeking to protect AI-related inventions in China. On one hand, it appears that the. CNIPA maintains a close watch against attempts to claim generic algorithms without limiting to certain specific application field(s) to sufficiently distinguish from known art. On the other hand, if an applicant can pinpoint a specific technical problem in a particular application field, provide tailored model training, and identify distinct output uses that result in surprisingly improved technical effects not disclosed in the known prior art, they may have a strong chance of establishing inventive step.

Under the three-step framework for assessing inventive step, the Board emphasized that examiners must consider whether the change of application field results in substantial adjustments or modifications to the algorithm or model. Specifically, where a new application field leads to significant changes in:

• the structure of the model,
• the training data it processes,
• the input or output data involved, and/or
• the selection and configuration of algorithms,

and such differences produce a non-obvious technical effect relative to prior art, the invention should be recognized as inventive.

The reasoning parallels the 2022 Top 10 case³ of “Method for establishing a neural network model for classifying grades of scrap steel”, where the inventive step was recognized based on the adaptation of a neural network model to a highly specialized industrial context.

For patent applicants and patentees of AI-related inventions, this case highlights some strategic considerations when drafting the applications:

• Describing the application field(s) in detail: If the invention has multiple potential applications, be explicit in describing these application fields/scenarios in the specification. Each scenario should define not only the input and output but also the technical problem unique to that context.
• Highlight training adaptation: If the algorithm itself is not novel, emphasize inventive aspects in the sourcing of data used for training, the object-specific tailoring, and the use of the output data to achieve the intended technical effect specific to the application scenario.
• Fallback positions: Consider including fallback positions in dependent claims or numbered embodiments limiting to one or more application fields to preserve opportunities to amend the claim scope during re-examination or invalidation proceedings.

Looking ahead, similar disputes will almost certainly arise as AI technologies continue to develop and overlap across industries. The lesson is clear: broad yet detailed examples, explicit technical problems, and clear linkages between model training and application scenarios are extremely important.

This article is for general informational purposes only and should not be considered legal advice or a legal opinion on a specific set of facts.

1. Sessions 6.1.3 and 6.2, Part II Chapter 9, Patent Examination Guidelines 2023

   ↩︎

2. https://www.cnipa.gov.cn/art/2025/6/6/art_2648_200004.html

   ↩︎
3. https://www.cnipa.gov.cn/art/2023/4/26/art_3207_184728.html ↩︎