Retrieval Model Training with Cross-Screened Triplets for PQ Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image retrieval methods using product quantization (PQ) suffer from issues such as missed recalls and false recalls due to similar features being split into adjacent segments, leading to inaccurate retrieval results.
Innovation Solution
A method involving an embedding vector network and a quantization index network is trained using sample triplets to improve noise recognition capabilities, where the embedding vector network screens out a first sample triplet set for training the quantization index network, and vice versa, enabling denoising learning and improved prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If product quantization is used to divide feature vector dimensions into segments, then retrieval speed is improved, but similar features are split into adjacent segments causing missed recalls and false recalls
Solution Approach 1:
The patent applies segmentation by dividing the feature vector into multiple sub-vectors, where each sub-vector corresponds to a specific segment. This allows the system to retrieve candidate images from multiple segments simultaneously rather than relying on a single segmented structure, thereby maintaining retrieval speed while reducing missed recalls caused by feature splitting.
Solution Approach 2:
The patent merges multiple retrieval results from different segments by combining candidate image sets from each segment. This integration approach ensures that images potentially missed in one segment are recovered from other segments, resolving the accuracy issue while preserving the speed benefits of segmented quantization.
2Reliability
If multiple candidate images are recalled for sorting, then retrieval accuracy is improved, but computational complexity and time consumption increase
Solution Approach 1:
The patent implements partial action by retrieving a predetermined number of candidate images from each segment rather than exhaustively searching all possible candidates. This selective approach achieves sufficient retrieval accuracy while significantly reducing computational complexity and processing time.
Solution Approach 2:
By segmenting the retrieval process into multiple independent segment queries, the system distributes computational load across parallel operations. This segmentation strategy reduces overall computational complexity while maintaining or improving accuracy through the aggregation of results from multiple segments.
3Reliability
If critical samples are recalled multiple times to avoid missed recalls, then retrieval completeness is improved, but false recalls increase
Solution Approach 1:
The patent employs feedback mechanisms by sorting and evaluating retrieved candidate images based on their relevance to the query. This feedback loop allows the system to identify and prioritize truly relevant images while filtering out false recalls, thereby maintaining retrieval completeness without increasing false positives.
Solution Approach 2:
The system retrieves more candidate images than strictly necessary (excessive action) from multiple segments, then applies sorting and filtering to select the final relevant results. This approach ensures completeness by capturing all potential matches while using subsequent filtering to eliminate false recalls.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method and apparatus for training a retrieval model, a retrieval method and apparatus, a device and a medium are provided, belonging to the field of artificial intelligence. The retrieval model includes an embedding vector network and a quantization index network. The foregoing method includes: obtaining n sample triplets used for training the retrieval model (320); inputting basic feature vectors of the n sample triplets to the embedding vector network; screening out, according to errors of feature vectors output by the embedding vector network, a first sample triplet set used for training the quantization index network (340); inputting the basic feature vectors of the n sample triplets to the quantization index network; screening out, according to errors of quantization indexes output by the quantization index network, a second sample triplet set used for training the embedding vector network (360); training the quantization index network on the basis of the first sample triplet set (381); and training the embedding vector network on the basis of the second sample triplet set (382). The above method improves the accuracy of a retrieval model.