T 1526/20 (Liveness testing/SAMSUNG) 14-03-2023

Reasons for the Decision

The application

1. The application relates to liveness testing in the framework of face recognition methods; it is in par­ticular concerned with the detection of impersonations using a picture of a face (paragraphs 1 and 2). The proposed method relies on the observation that the image of a 3D structure (i.e. of a real person) is less "uniform" than that of a 2D structure (of a picture of the person). To detect "uniformity" it applies first a diffusion filter and then measures the diffusion speeds, i.e. the magnitude of the change in pixel values before and after filtering (see e.g. paragraphs 7 to 14, 102, 103). These "speeds" are binarized and if the number of non-zero pixels is low, then the imaged object is considered to be a fake (paragraphs 108-117). Alternatively, the binarized image is median filtered, and if the number of (by the filter) changed pixels is low, then the imaged object is considered to be a fake, on the basis that the image is not noisy enough to be one of a real object (paragraphs 118-119).

2. The application further describes (paragraph 143 et seq.) a method for face recognition, also based on diffusion, which includes liveness testing as previously described. In that context it is explained (see e.g. paragraphs 163-164, 194-195) that the diffused image contains (mostly) the illumination component of the given image. Another image is formed by the ratio, or equivalently the log difference, of the original image to the diffused image. This image is said to be related to the "non-illumination component", i.e. to the structure of the imaged object, and is the one used for recognition.

The cited documents

3. Document D1 describes liveness testing by decomposing the input face image into a reflectance image (albedo) and an illuminance image, using the Logarithmic Total Variation (LTV) method of D3 (section 2.2). These images are used either separately, or together, as features in a classifier, e.g. the sparse logistic classifier (section 2.3).

4. Document D2 is not prior art, having been published af­ter the priority date. It is a paper authored by three of the designated inventors and describes a method of liveness testing using diffusion, which is in concept similar to (see section II.A), but in implementation different from, the one described in the current appli­cation. In particular, the diffusion speed is computed using the log difference (II.B), and the feature extrac­tion step employs a concatenated histogram of local speed patterns (II.C).

Second auxiliary request

Correction, Rule 139 EPC

5. The second auxiliary request filed with the grounds of appeal still contained the dependent claim 5, to which the Examining Division had objected under various pro­vi­sions, although the Appellant declared to have dele­ted it (statement of grounds of appeal, page 9, middle paragraph) like in its other requests. The Appellant also did not submit any arguments against the corres­pon­ding grounds for refusal.

6. In its provisional opinion, the Board indicated its understanding that the intention of the Appellant was to delete this claim and asked the Appellant to clarify its requests. In response, the Appellant filed the current second auxiliary request as a "corrected" ver­sion, which differs from the previous second auxiliary request only by the deletion of claim 5. The Board accepts that this reflects the Appellant's original intention and considers the corresponding amendment as an allowable correction under Rule 139 EPC.

Fourth auxiliary request

Admittance, Article 13 RPBA 2020

7. The fourth auxiliary request was filed in response to the Board's preliminary opinion. It addresses, and overcomes, the lack of clarity objection raised by the Board (for the first time) in its preliminary opinion at point 7.2, without raising any other issues. In view of this, the Board decides to admit the amended request (see also T 1294/16 reasons 8.2 to 8.4).

Main and first to fourth auxiliary requests

Sufficiency of disclosure, Article 83 EPC

8. In the oral proceedings the Board indicated that it would first examine the fourth auxiliary request, which defined the term "diffusion speed" in a clear manner. The Appellant had no objections.

9. In its decision, the Examining Division objected to the independent claims of all requests (see especially the decision, point 7.1 of the reasons) because the claimed techniques, "relying only on diffusion filtering" were not robust against illumination changes and lead to "to­tally unreliable results". For this reason other tech­niques, as presented in D1 or D2, had to include further steps. A similar objection was made in view of the computation of the diffusion speeds (decision, point 7.2), which in D2 (as well as in D1 and D3, though they do not use the term "diffusion speed") was made in logarithmic space, so as to obtain consistent results.

10. The Appellant argued (see for instance the grounds of appeal, pages 5 to 7) that the invention might not work in some hand-picked examples, but worked in "the majority of practical use cases". The approach was not dependent on the illumination used, because it was the shape of the surface considered that determined the diffusion speeds.

10.1 The logarithmic differences provided robustness to changes in illumination for face recognition, but were not necessary for liveness testing. The Appellant furt­her emphasised this distinction during the oral procee­dings. Face recognition was a more complex task than liveness testing, which could be carried out based only on the normal difference of the two images. The Exami­ning Division did not sufficiently appreciate this dis­tinction. With reference to figure 12, the Appellant explained that liveness testing did not need the compu­ta­tion of the third image, i.e. the logarithmic diffe­rence image.

10.2 The skilled person, at the time of filing, had no rea­sons to doubt that the proposed method worked. Even if the authors later used the logarithmic space for diffe­rencing, the diffusion speeds computed in the "Carte­sian" space were also sufficient to determine liveness.

10.3 This could be seen from section 2.2 in document D1. Although a logarithmic difference was used in the part using diffusion filtering, entitled "Variational Retinex-based Method", this was not the case in the following part, entitled "Difference of Gaussian (DoG)-based Method": There, a DoG filter was used, which implemented a normal difference. The idea expressed in that part, namely that because

"the image of a photograph taken through a webcam is essentially an image of a real face but passes through the camera system twice and the printing system once [...] the imposter image tends to be more seriously distorted by the imaging system and hence has lower image quality (i.e., missing more high frequency details) under the same imaging conditions"

was the same as that of the current application (see paragraphs 76 and 77). Essentially, high frequency patterns, i.e. referred to as non-uniform in the pre­sent application, could be used to distinguish a live image, i.e. an image of a live face, from a fake one, i.e. the image of a 2D photograph of a face: due to its 3D profile, the live image contained more, (and diffe­rent) high frequency information than the image of a 2D photograph.

10.4 The application disclosed a clear sequence of steps which the skilled person could implement without undue burden. It explained that, in an image of a real face, small scale regions with a pronounced 3D structure had higher diffusion speeds than large and flat scale regions such as cheeks (paragraph 105). So a live image could be distinguished from a fake one based on the diffusion speeds using equations 6 or 7. In particular the embodiment using median filtering (equation 7) was a robust implementation of this idea and made it credible to the skilled person that the method could provide for reliable liveness testing.

10.5 The Appellant also argued that the sufficiency of dis­closure of an invention in a patent application should not be assessed according to the stricter criteria applicable to a scientific publication.

11. The Board notes that the application, as the Appellant also submitted, discusses two distinct scena­rios, the first relating to liveness detection and the second to face recognition. In both scenarios, the first step is to apply image diffusion to the original image.

11.1 For face recognition, the application provides a multiplicative image formation model: the acquired image is said to be the product between illumination and object information (reflectance). According to the application, the illumination information (second image) can be extracted by diffusion filtering. The object-related information (third image) is then extracted by a log-difference between the original image and the diffused image - or, equivalently, the logarithm of the ratio of the first two images , i.e. by taking the logarithm of the object-related information according to the multiplicative model.

12. The Board also notes that this multiplicative model is well known to the skilled person as the Lambertian reflectance model for image formation (see e.g. D1 equation 1) and is used in D1, section 2.2, to justify the use of the same log-difference, namely between the original image and a diffused image, to obtain an image used for liveness testing (the albedo image).

13. The application uses the "diffusion speeds" for live­ness testing. They constitute a (third) image obtained by a (normal) difference between the original and the diffused image (see the application, equation 4). This rather corresponds to an additive model for image forma­­tion, in apparent contradiction with the generally accepted multiplicative theoretical model. The applica­tion contains nothing to dismiss that contradiction. Therefore, on the basis of the application and the common general knowledge, which does not include D2, the skilled person must have had doubts that this model successfully extracted object-related information and, thus, provided the information necessary for reliable liveness testing.

14. The application provides no reasons as to why this model was correct, nor results to show that the proposed way of extracting object-related information correctly distinguishes live from fake objects.

14.1 That a detailed implementation method is provided, as the Appellant pointed out, does not withstand this objection, as the implementation cannot establish that the claimed effect exists.

14.2 Beyond that, neither the allegation in the application that the diffusion speeds would be characteristic for 3D "small scale" regions and thus contain the relevant discriminatory information nor the similar argument reproduced from D1 (point 12.3 above) can be accepted without further evidence.

14.3 This is because the purpose requires information distinguishing between images obtained from 3D and 2D structures. Although some filters may be able to extract such information (as it was shown for the DoG filter in D1), a generic low pass filter, (e.g. an average filter) or a generic diffusion filter cannot be presumed to do the same, as they are designed for a different purpose, namely general denoising. Depending on the filter and its parameters, the identified "noise" may or may not capture the needed information. Additional evidence is required to show which image acquisition conditions und which filters and filter parameters may be suitable for the task at hand. This is missing in the application at hand.

14.4 The Board concludes that the application does not dis­pel the skilled person's doubts that any one of the provided implementation examples is able to determine liveness in a reliable manner. The claimed invention is therefore not sufficiently disclosed in the sense of Article 83 EPC.

15. This conclusion applies a fortiori to the main and first to third auxiliary requests as well, which are effectively broader in scope. The Appellant did not provide any counter-arguments in this respect.

Fifth auxiliary request

Admittance, Article 13 RPBA 2020

16. This request details on the basis of equation 7 the manner in which the discrimination between a fake and a live image takes place. The Appellant argued that the skilled person, in view of the explanations in paragraphs 76 and 77 of the description, would find credible that the theoretical model so defined worked. The claim now provided a clear teaching of the steps to be executed, which could be implemented without any difficulties by the skilled person.

17. The Board notes that the input to the discrimination method is still the "diffusion speed" image, obtained as the difference between the original image and the diffused one (n.b. the specification "nonlinear diffu­si­vity function-based" does not effectively restrict the claim, because all image diffusion methods, as defined by equation 1, are non-linear). As noted above, the Appellant could not dispel the doubts that this image contains the information necessary for liveness discrimination.

18. Thus the request, filed very late during the oral pro­ceedings, does not overcome the objection as to a lack of sufficient disclosure. The Board therefore decided not to admit this request (Article 13 RPBA 2020).