Image Search Re-Ranking for Multi-Camera Person Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image search systems struggle to efficiently identify the same person across images captured from multiple cameras due to similar features among individuals and the computational burden of calculating similarity metrics like Jaccard distances.
Innovation Solution
Implement a method that performs primary ranking based on image similarity and re-ranking using Jaccard distances, while limiting the search target size by filtering based on metadata such as sex, capture time, and location, utilizing a trained neural network to extract feature vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If feature information extracted by deep learning network is used to search for the same person, then search capability is improved, but reliability deteriorates when multiple people have similar features
Solution Approach 1:
The patent implements pseudo relevance feedback by using the initial search results to generate query expansions and refine the search. The system automatically analyzes the initial results, extracts relevant visual features, and uses these to re-rank results, creating a feedback loop that improves reliability without requiring manual intervention.
Solution Approach 2:
The patent changes the parameter of similarity measurement from basic feature matching to Jaccard distance calculation on visual features. This parameter change allows the system to better distinguish between different people with similar features by considering the overlap and uniqueness of visual characteristics.
2Measurement precision
If Jaccard distance calculation is performed on all candidate images, then search accuracy is improved, but computational load increases rapidly
Solution Approach 1:
The patent segments the search process into two stages: initial filtering using basic feature matching, and refined re-ranking using Jaccard distance. This segmentation allows the computationally intensive Jaccard calculation to be applied only to a small subset of candidate images rather than the entire database.
Solution Approach 2:
The patent applies partial action by performing Jaccard distance calculation on only the top-k candidate images from the initial search results, rather than on all images in the database. This partial application of the precise but computationally expensive operation achieves high accuracy while controlling computational load.
3Adaptability or versatility
If vast amounts of image data are stored for comprehensive search, then search coverage is improved, but search efficiency deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-extracting and storing visual features for all images in the database. This pre-processing allows the search system to quickly retrieve and compare features without performing computationally intensive feature extraction during the actual search operation.
Solution Approach 2:
The patent introduces visual feature vectors as an intermediary representation between raw images and search results. These feature vectors serve as a compact, searchable representation that enables efficient comparison and matching while preserving the essential visual information needed for accurate person identification.
Data Source
AI summary
A device and a method for image search in a surveillance camera system are disclosed. An image search device according to an embodiment of the present specification comprises a database for storing images captured from a plurality of cameras spaced apart from each other and metadata of the images, wherein, when a probe image is input, the device restricts a search range for searching for the same person as a person of interest in the probe image according to a predetermined criterion, performs primary ranking on an image within the restricted search range on the basis of feature vector information of the image, and then performs re-ranking according to an additional image similarity on the basis of a result of the primary ranking. Accordingly, the accuracy of an image search can be improved while reducing the amount of computation required for the image search. In the present specification, one or more among a surveillance camera, an autonomous vehicle, a user terminal, and a server may be linked to an artificial intelligence module, a robot, an augmented reality (AR) device, a virtual reality (VR) device, a device related to 5G services, and the like.