Facial Recognition via DCT Features and Pupil Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition technologies face challenges in reliability due to variations in lighting conditions and normalization errors, particularly with methods like eigenface and 3D models, which are either expensive or prone to recognition errors from node correspondence issues in generalized face bunch graphs.
Innovation Solution
The method involves normalizing facial images by detecting eye pupil centers, rotating and scaling them to fixed coordinates, and using Discrete Cosine Transformation (DCT) features to calculate distances between input and template images, with optional 3D stereo-pair reconstruction for robustness, and employing histogram equalization or local contrast normalization to reduce lighting effects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If eigenface technique with KL basis decomposition is used, then image recognition can be performed, but recognition reliability deteriorates under lighting variations and normalization errors
Solution Approach 1:
The patent transforms the facial image representation from direct pixel values to DCT coefficient space, changing the parameter domain to achieve illumination invariance. The DCT transformation converts spatial domain variations (caused by lighting) into frequency domain representations where illumination effects can be separated and removed, thereby improving recognition reliability without sacrificing implementation simplicity
Solution Approach 2:
The patent introduces DCT coefficients as an intermediary representation between the raw facial image and the recognition decision. This intermediary transforms the input data into a form where lighting variations are decoupled from identity features, allowing the system to maintain high reliability across different lighting conditions while keeping the overall system architecture simple
2Reliability
If 3D models of faces composed of 3D models of face elements are used, then recognition can be performed, but cost increases due to requirement of laser range finders
Solution Approach 1:
The patent replaces expensive, complex 3D reconstruction hardware (laser range finders) with a simpler, computationally-based approach using standard 2D image capture and DCT processing. This substitution achieves comparable or superior recognition accuracy through software-based illumination removal rather than hardware-based 3D modeling, dramatically reducing system cost while maintaining reliability
Solution Approach 2:
The patent substitutes the mechanical/optical 3D reconstruction system with a computational image processing system. Instead of using physical laser range finders to capture 3D geometry, the system uses 2D images processed through DCT transformation to achieve illumination-invariant recognition, replacing complex hardware with efficient software algorithms
3Difficulty of detecting and measuring
If generalized face bunch graph with Gabor filters is used, then face features can be extracted, but recognition accuracy deteriorates due to errors in node correspondence detection
Solution Approach 1:
The patent extracts only the essential frequency components through DCT transformation, removing the problematic aspect-based decomposition of generalized face bunch graphs. By working directly with DCT coefficients that represent global frequency patterns rather than local Gabor filter responses, the system eliminates node correspondence errors while retaining the ability to extract meaningful face features
Solution Approach 2:
The patent inverts the traditional approach of detecting local features first and then matching them. Instead, it applies global DCT transformation to the entire image first, then extracts features from the frequency domain representation. This reversal of the processing sequence eliminates the need for complex node correspondence detection while maintaining feature extraction effectiveness
4Device complexity
If not all components of Gabor filter feature vectors are informative, then computation is simplified, but recognition accuracy decreases due to decreased distance between different faces
Solution Approach 1:
The patent changes the feature representation from Gabor filter coefficients to DCT coefficients, transforming the parameter space to one where illumination effects are naturally separated. This parameter change allows the system to use a compact set of DCT coefficients that are inherently more discriminative, achieving both computational efficiency and high recognition accuracy simultaneously
Data Source
AI summary
A method to recognize a facial image is described. An input facial image is normalized by scaling and rotation angle using methods of eye pupil centers detection. The input facial image is further normalized by lighting intensity. Template images are obtained either by the processing of certain images taken from different face positions, or by a preliminary reconstruction of a 3D face model based on stereo-pair images. Using the 3D model template, facial images are generated at different rotation angles. Distances between the input facial image and the template image are calculated from the Discrete Cosine Transformation (DCT) features, defined by overlapped blocks of these images. The facial image is recognized based on these distances.