Retrieval Model Training With Triplet Denoising for PQ Recall Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image retrieval methods using product quantization (PQ) often misclassify or miss similar features due to segmentation issues, leading to incorrect or incomplete recall of candidate objects.
Innovation Solution
A retrieval model is trained using an embedding vector network and a quantization index network, where sample triplets are used to screen and train both networks to improve noise recognition, ensuring accurate prediction of positive and negative samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If product quantization (PQ) is used to segment feature vectors into multiple segments, then the retrieval process becomes faster and more efficient, but similar features are easily split into adjacent segments causing critical samples to be missed or repeatedly recalled
Solution Approach 1:
The patent applies segmentation by dividing the feature vector space into multiple quantization segments, where each segment is represented by a numerical code. This allows efficient retrieval by grouping similar features while maintaining the ability to distinguish critical samples through learned segment boundaries that prevent oversplitting of similar features.
Solution Approach 2:
The patent changes the parameters of the quantization process by learning optimal segment boundaries and representations through training with sample triplets. This transforms the fixed PQ approach into an adaptive system that adjusts segmentation parameters to preserve feature similarity while maintaining retrieval efficiency.
2Device complexity
If traditional product quantization is used, then the system complexity remains low, but the retrieval accuracy deteriorates due to easy splitting of similar features into adjacent segments
Solution Approach 1:
The patent introduces an intermediary training mechanism using sample triplets that mediates between the simple PQ structure and the need for high accuracy. The training process acts as an intermediary step that optimizes the quantization parameters without fundamentally complicating the underlying PQ architecture.
Solution Approach 2:
The patent applies preliminary action by pre-training the quantization index network and embedding vector network using carefully constructed sample triplets before actual retrieval. This preliminary training phase prepares the system to handle critical samples correctly while maintaining the simplicity of the PQ structure during runtime.
3Reliability
If more candidate objects are recalled to ensure high confidence results, then the retrieval completeness improves, but the number of false recalls increases
Solution Approach 1:
The patent implements feedback through the use of sample triplets during training, where positive samples (similar features) and negative samples (dissimilar features) provide feedback signals that guide the optimization of the embedding vector network and quantization index network. This feedback mechanism enables the system to learn to distinguish true positives from false positives.
Solution Approach 2:
The patent combines multiple components into a composite retrieval system that integrates the embedding vector network, quantization index network, and traditional PQ methodology. This composite approach leverages the strengths of each component to achieve both high completeness and low false positive rates.
Data Source
AI summary
A method for training a retrieval model is performed by a computer device. The retrieval model includes an embedding vector network and a quantization index network. The method includes: obtaining n sample triplets; inputting basic feature vectors of the n sample triplets to the embedding vector network to identify, according to errors of feature vectors output by the embedding vector network, a first sample triplet set used for training the quantization index network; inputting the basic feature vectors of the n sample triplets to the quantization index network to identify, according to errors of quantization indexes output by the quantization index network, a second sample triplet set used for training the embedding vector network; training the quantization index network on the basis of the first sample triplet set; and training the embedding vector network on the basis of the second sample triplet set.