PTAB

IPR2018-00688

Nikon Corporation v. Carl Zeiss AG

Log In

1. Case Identification

2. Patent Overview

  • Title: Image Pattern Detection and Recognition
  • Brief Description: The ’163 patent discloses methods and systems for detecting image patterns, particularly human faces. The core invention is a two-step "pre-filtering" technique to efficiently identify candidate regions in an image by first correlating the image with reference data and then screening the resulting regions based on a grayscale characteristic.

3. Grounds for Unpatentability

Ground 1: Obviousness of Claims 1-3, 6, 9-10, 14-16, and 19 over Yow

  • Prior Art Relied Upon: Yow (Image & Vision Computing 15 (1997)).
  • Core Argument for this Ground:
    • Prior Art Mapping: Petitioner argued that Yow discloses all limitations of the independent claims. Yow teaches a two-stage "preattentive" feature selection for face detection that mirrors the ’163 patent's pre-filtering method. Yow’s first step performs "spatial filtering" using a matched bandpass filter (a Gaussian derivative) to find "local maxima," which Petitioner asserted is equivalent to the ’163 patent's step of correlating an image to derive a correlation image and finding local maxima. Yow’s second step examines the identified interest points by performing edge detection and analyzing image characteristics like "gray-level variance" to screen for valid facial features, which corresponds to the ’163 patent's screening step based on a grayscale characteristic.
    • Key Aspects: The argument centers on the direct mapping of Yow’s two-step "preattentive feature selection" to the ’163 patent's two-step "pre-filtering" method, contending that both use a correlation-like step followed by a grayscale-based screening step to reduce data for a final face detector.

Ground 2: Obviousness of Claims 4, 7, and 11 over Yow in view of Crowley

  • Prior Art Relied Upon: Yow (Image & Vision Computing 15 (1997)) and Crowley (Robotics and Autonomous Systems 16 (1995)).
  • Core Argument for this Ground:
    • Prior Art Mapping: This ground addressed dependent claims requiring the use of training images to create the reference data (claim 4) and thresholding of correlation values (claims 7 and 11). While Yow provided the basic two-step framework, Petitioner argued Crowley explicitly taught these missing elements. Crowley, a well-known reference in face detection, disclosed deriving a convolution kernel by averaging a plurality of training images. Crowley also taught comparing correlation values to a threshold value to select face candidates.
    • Motivation to Combine: A person of ordinary skill in the art (POSITA) would combine the references because both addressed face detection. Yow suggested using various filters, and Crowley’s method of creating a filter from training images was a well-known technique to improve robustness. A POSITA would be motivated to incorporate Crowley's superior filter generation and thresholding techniques into Yow’s pre-filtering framework to achieve a more accurate and efficient system.
    • Expectation of Success: The combination was presented as a predictable integration of known techniques. Replacing Yow's filter with one derived using Crowley's method and adding a standard thresholding step were argued to be straightforward modifications that would predictably result in improved face detection.

Ground 3: Obviousness of Claims 1, 4, and 6-7 over Kosugi

  • Prior Art Relied Upon: Kosugi (Japanese Patent Pub. Hei 5-174149).

  • Core Argument for this Ground:

    • Prior Art Mapping: Petitioner asserted that Kosugi, an image recognition system, disclosed a multi-stage process that constituted the claimed pre-filtering method. Kosugi’s first "Rough Search" stage correlated a roughly pixelated input image with a target by calculating distances in intensity space, which Petitioner contended was analogous to creating a correlation image and finding local maxima (minimum distance). Kosugi's second "Detailed Search" stage further screened candidate regions by analyzing finer grayscale characteristics and discarding regions that did not meet a criterion, mapping to the ’163 patent's second step. Kosugi also explicitly taught creating reference data from a plurality of face images.
    • Key Aspects: Petitioner argued that although Kosugi calculated "distance" instead of "correlation," a POSITA would understand that minimum distance corresponds to maximum correlation. Creating a formal correlation image from Kosugi's distance matrix was argued to be an obvious design choice.

  • Additional Grounds: Petitioner asserted additional obviousness challenges, including grounds based on Yow’s Ph.D. thesis (which largely mirrors the Yow article). Further combinations included adding Fang (WO 96/38808) to Kosugi to teach comparing contrast values to a threshold, and further adding Rowley (an applicant-admitted reference) to teach using a neural network as the final image pattern detector for system claims 15 and 19.

4. Key Claim Construction Positions

  • "Correlating" / "Correlation": Petitioner proposed construing these terms to mean "an image correlation operation between the input image and a kernel that results in a correlation image." This construction was argued to be critical, as it requires the output to be a two-dimensional image of the same size as the input. Petitioner supported this by citing the patent’s specification and prosecution history, where the patentee allegedly distinguished prior art for failing to produce a "correlation image."
  • "Convolution": Petitioner argued this term should be construed identically to "correlation," as the specification used the terms interchangeably (e.g., "The linear convolution step, also called 'linear correlation'…").
  • Means-Plus-Function Terms (Claim 15): Petitioner contended that "first filtering means," "second filtering means," and "image pattern detector" were means-plus-function terms under §112, ¶6.
    • "first filtering means": The function was identified as correlating an image to select regions, and the structure was identified as the "linear matched filter 56" programmed with a specific algorithm.
    • "second filtering means": The function was identified as screening regions based on a grayscale characteristic, and the structure was the "non-linear filter 66" programmed to analyze contrast values.
    • "image pattern detector": The function was to verify if a candidate region contains the target pattern, and the structure was limited to the single disclosed embodiment: the "neural-network face detection system."

5. Relief Requested

  • Petitioner requested institution of an inter partes review and cancellation of claims 1-4, 6-7, 9-11, 14-16, and 19 of Patent 6,463,163 as unpatentable.