Multi-Image Clothing Recognition via Frequency Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image evaluation systems face challenges in accurately detecting clothing regions in images due to changes in pose and illumination, leading to incorrect recognition and evaluation of clothing types, especially when the pose of the person changes or the direction they are facing.
Innovation Solution
An image evaluation apparatus that specifies clothing types for each person in images and characterizes image groups based on the frequency of appearance of different clothing types, allowing for correct evaluation even when individual image recognition is incorrect by referencing correctly recognized clothing types in other images within the group.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If image characteristics are extracted from detected clothing regions for recognition, then clothing type recognition can be performed, but recognition accuracy deteriorates when pose or illumination changes
Solution Approach 1:
The patent transitions from analyzing individual images in isolation to analyzing multiple images collectively. By extracting clothing region information from multiple images showing the same person and aggregating this information, the system moves from a 2D single-image analysis to a multi-dimensional approach that incorporates temporal and contextual dimensions, thereby improving recognition accuracy despite pose or illumination variations in individual images
Solution Approach 2:
The patent merges clothing region information from multiple images into unified clothing information for each person. By combining data from several images showing the same person in different poses or lighting conditions, the system creates a more robust and accurate representation of the clothing being worn, overcoming the limitations of any single image
2Productivity
If clothing recognition is performed on individual images, then processing speed is maintained, but evaluation accuracy deteriorates when recognition is incorrect
Solution Approach 1:
The patent combines clothing recognition results from multiple images processed in parallel, maintaining processing efficiency while improving reliability. By aggregating information from multiple images and using voting or frequency-based methods to determine the most likely clothing type, the system achieves both high productivity and high reliability
3Device complexity
If clothing regions are detected in each image independently, then processing complexity is reduced, but recognition reliability deteriorates due to pose and illumination variations
Solution Approach 1:
The patent performs preliminary detection of clothing regions in multiple images independently (maintaining low complexity), then uses this preliminary information as input for a subsequent aggregation step. This two-stage approach allows the system to maintain simple, efficient detection algorithms while improving reliability through collective analysis of multiple images
Data Source
AI summary
In an image evaluation apparatus, a clothing recognition unit performs, for each person appearing in each of images included in an image group generated by an image group generation unit, recognition of clothing that the person is wearing. An image event evaluation unit, according to types of clothing recognized by the clothing recognition unit and a frequency of appearance of each type of clothing in the images in the image group, collectively evaluates the images included in the image group.