Face Image Grouping Using Statistical Similarity Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for creating image products with faces are time-consuming and lack accuracy in grouping and selecting faces, especially when dealing with multiple individuals, as they often require prior training and supervision.
Innovation Solution
A computer-implemented method that automatically categorizes face images using probabilistic descriptions and similarity functions, allowing for the grouping of faces without prior knowledge or supervision, and utilizes non-negative matrix factorization to improve the accuracy of face grouping in image products.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional face detection techniques are used that assume single-face regions are already detected, then face recognition accuracy may be maintained, but the system cannot properly group multiple faces of the same person across different images
Solution Approach 1:
The patent segments the face recognition task into two distinct stages: first, detecting individual face regions in each image using conventional single-face detectors; second, grouping detected faces across multiple images by comparing facial features to identify matches. This segmentation allows the system to leverage existing single-face detection accuracy while adding multi-face grouping capability through systematic comparison of detected faces across the image set.
Solution Approach 2:
The patent performs preliminary face detection and extraction of facial features from each image before attempting to group them. By pre-processing the images to identify and extract face regions and their characteristic features, the system creates a structured set of candidate faces that can then be efficiently compared and grouped, rather than attempting to group raw images directly.
2Measurement precision
If training-based face recognition methods are used with known person data, then identification accuracy improves, but the system requires prior supervision and cannot handle unknown individuals
Solution Approach 1:
The patent implements a self-service face grouping system that does not require pre-training with known person data. Instead, the system automatically extracts facial features from detected faces and performs unsupervised grouping by comparing features across images, using statistical methods to identify matches without human supervision or prior knowledge of the individuals in the images.
Solution Approach 2:
The patent changes the operational parameters from supervised classification (requiring labeled training data) to unsupervised feature-based clustering. By transforming the problem from identifying known persons to grouping faces based on feature similarity, the system achieves adaptability to unknown individuals while maintaining grouping accuracy through statistical comparison methods.
3Quantity of substance
If multiple pictures of the same person at the same scene are included, then comprehensive coverage is achieved, but the image product quality deteriorates due to redundancy
Solution Approach 1:
The patent extracts and compares key facial features from each detected face to identify redundant pictures showing the same person at the same scene. By extracting distinctive facial characteristics and systematically comparing them across the image set, the system identifies and removes duplicate representations, keeping only one picture per unique person-scene combination while maintaining comprehensive coverage of all individuals.
4Productivity
If automated face grouping without prior training is implemented, then processing time is reduced, but grouping accuracy may worsen
Solution Approach 1:
The patent performs preliminary extraction and normalization of facial features from detected faces before grouping. By pre-processing the facial data to extract stable, comparable characteristics and organizing them in a standardized format, the system enables rapid unsupervised grouping without sacrificing accuracy, as the feature extraction creates a robust basis for comparison that works well without training data.
Solution Approach 2:
The patent changes the grouping approach from training-intensive supervised learning to efficient unsupervised feature matching. By transforming the input data into a standardized feature space and using statistical comparison methods, the system achieves fast processing speeds while maintaining grouping accuracy through the quality of feature extraction and comparison algorithms rather than through training data volume.
Data Source
AI summary
A computer-implemented method for creating an image product by accurately grouping faces includes receiving an initial set of face groups for a plurality of face images, training classifiers between pairs of face groups in the initial set of face groups using image-product statistics by a computer processor, classifying the plurality of face images by classifiers to output binary vectors for the plurality of face images by the computer processor, calculating a value for an improved similarity function using the binary vectors for each pair of the plurality of face images, grouping the plurality of face images into modified face groups based on values of the binary similarity functions by the computer processor, and creating an image product based at least in part on the modified face groups.