Face Feature Registration via Grouped Selection for Database Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face collation techniques in surveillance videos face challenges in managing large image feature amounts, leading to database bloat and reduced accuracy due to frame thinning and extensive registration processes.
Innovation Solution
An information processing apparatus that detects faces, classifies them into groups, selects a predetermined number of faces per group, and registers only the most relevant feature amounts, prioritizing frontal face directions for higher accuracy and reduced storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all faces detected from all frames of a video are registered, then the collation accuracy is improved, but the number of image feature amounts to be stored becomes huge
Solution Approach 1:
The patent segments faces into multiple groups based on detection order and characteristics, then selects only a predetermined number of representative faces from each group for registration. This segmentation approach maintains collation accuracy by ensuring diverse face representations while dramatically reducing the total number of feature amounts stored compared to registering all detected faces.
Solution Approach 2:
The patent extracts and selects only the most representative faces from each group based on specific criteria (detection order, face characteristics), discarding redundant faces. This extraction principle directly reduces the quantity of stored feature amounts while preserving the essential information needed for accurate collation.
2Quantity of substance
If frame thinning is applied to reduce the number of image feature amounts, then the storage requirement is reduced, but the accuracy of face collation deteriorates due to discarding frames with favorable face directions
Solution Approach 1:
The patent implements a feedback mechanism where faces are classified into groups based on detection order and characteristics, and the selection of representative faces considers the distribution across different frames and face directions. This feedback-based selection ensures that important face instances are retained while still reducing the overall number of stored features, avoiding the accuracy loss associated with simple frame thinning.
Solution Approach 2:
The patent changes the selection criterion from simple frame thinning (time-based reduction) to a more sophisticated selection based on face group characteristics and detection order. This parameter change allows the system to maintain collation accuracy by preserving faces with favorable characteristics while still achieving storage reduction through selective registration.
3Productivity
If faces are classified into multiple groups and only a predetermined number are selected per group, then the registration load is reduced, but the risk of missing favorable face directions increases
Solution Approach 1:
The patent performs preliminary classification of faces into groups based on detection order and characteristics before selecting representative faces. This preliminary action organizes the face data in a way that facilitates efficient selection while ensuring coverage of diverse face directions and characteristics, thereby maintaining collation accuracy despite selecting only a limited number of faces per group.
Solution Approach 2:
The patent applies different selection criteria and priorities to different face groups based on their local characteristics (detection order, face attributes). This local quality approach ensures that each group contributes its most representative faces to the registration, optimizing the balance between registration efficiency and maintaining comprehensive face direction coverage for accurate collation.
Data Source
AI summary
The present invention suppresses feature amounts of faces to be registered in a collation database, thereby preventing the collation database from being bloated. An information processing apparatus of the present invention includes an acquisition unit configured to acquire a video, a detection unit configured to detect at least one face of the same person from a plurality of frames of the acquired video, a classification unit configured to classify the detected faces into a plurality of predetermined groups, a selection unit configured to select, from the faces classified into the groups, not more than a first predetermined number of faces for each group, wherein the first predetermined number is an integer of not less than 2, and a registration unit configured to register feature amounts of the selected faces in a database.