Image Sample Selection via Information Quantum Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The selection of image samples for labeling in instance segmentation models is challenging, leading to poor training effects even with experienced labelers, as the quality of selected samples significantly impacts the model's performance.
Innovation Solution
A method involving the construction and training of instance segmentation and score prediction models using feature pyramid networks and region generation networks, followed by clustering and selection of target image samples based on information quantum scores and feature vectors, to improve the accuracy of sample selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If image samples are selected randomly or without careful selection, then the labeling process can be completed quickly, but the training effect of theinstance segmentation model will be poor
Solution Approach 1:
The patent applies preliminary action by pre-training an instance segmentation model and pre-computing feature vectors and information quantum scores for image samples before the actual labeling task. The score prediction model is trained in advance on labeled data to evaluate image quality. This preliminary preparation enables rapid selection of high-quality samples during the labeling process without requiring time-consuming real-time evaluation, thus resolving the contradiction between training effect reliability and time loss.
Solution Approach 2:
The system implements self-service by using the pre-trained instance segmentation model and score prediction model to automatically evaluate and rank image samples based on their information quantum scores and feature vector similarities. The clustering module automatically groups samples without human intervention. This automation eliminates the need for manual sample selection, allowing the system to self-evaluate and self-organize image data, thereby improving training effect while reducing time expenditure on sample selection.
2Reliability
If more image samples are selected for labeling, then the training effect can be improved, but the cost and time consumption increase
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different image samples based on their individual characteristics. Instead of uniformly processing all samples, the system evaluates each sample's information quantum score and feature vector to identify locally optimal high-quality samples. The clustering module groups samples with similar characteristics, allowing selective labeling of only the most informative samples from each cluster. This targeted approach improves training effect by focusing on high-value samples while maintaining labeling efficiency.
Solution Approach 2:
The system utilizes parameter changes by dynamically adjusting the number and selection criteria of labeled samples based on the trained model's performance and the distribution of information quantum scores. The score prediction model provides a quantitative parameter (information quantum score) that can be thresholded or ranked to determine the optimal subset of samples for labeling. This parameter-based selection allows flexible optimization of the balance between training effect and labeling cost, enabling the system to adapt to different project requirements.
3Measurement precision
If experienced labelers are used, then the labeling quality can be maintained, but the selection of appropriate image samples remains problematic
Solution Approach 1:
The patent introduces an intermediary system consisting of the score prediction model, feature vector extraction module, and clustering module that mediates between the raw image samples and the final labeled output. This intermediary automatically assesses sample quality by computing information quantum scores and grouping samples into clusters based on feature vector similarities. This mediation relieves experienced labelers from the difficult task of manual sample quality assessment, allowing them to focus solely on accurate labeling while the system handles the complex evaluation and selection process.
Data Source
AI summary
The present disclosure relates to a technology field of artificial intelligence and provides a method for selecting image samples and related equipment. The method trains an instance segmentation model with first image samples and trains a score prediction model with third image samples. An information quantum score of second image samples is calculated through the score prediction model and feature vectors extracted. The second image samples are clustered according to the feature vectors of the second image samples and sample clusters of the second image samples are obtained. Target image samples are selected from the second image samples according to the information quantum score of the second image samples and the sample clusters. Target image samples from the image samples are selected for labelling, improving an accuracy of sample selection.