Online Radiation Image Recognition Training With Partial Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional manual identification of radiation images is inefficient and lacks flexibility in adapting to changing user needs, with long implementation cycles and high manpower requirements for labeling all categories of interest, and existing systems fail to adjust to varying data distributions across sites.
Innovation Solution
A method for training a radiation image recognition model online by collecting and labeling images on-site to form incomplete training data sets, synthesizing with pre-stored data to create a training dataset, and training the model without returning data off-site, allowing for flexible category updates and reduced labeling burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional manual identification method is used, then data confidentiality is ensured, but implementation cycle is long and efficiency is low
Solution Approach 1:
The system enables on-site model training and updates without requiring external developer intervention. The radiation inspection system itself performs data labeling, model training, and deployment locally, making the system self-sufficient and eliminating the need to return data off-site for updates.
Solution Approach 2:
The system pre-stores preset data including first radiation images with complete labeling information locally. When new recognition needs arise, the system combines this pre-stored data with newly collected on-site data to form training datasets, enabling rapid model updates without external assistance.
2Manufacturing precision
If all objects of all categories are labeled manually, then complete training data is obtained, but labeling workload increases and labeling efficiency decreases
Solution Approach 1:
The system accepts incomplete labeling information from on-site collected data. Instead of requiring all objects to be labeled, it uses the available partial labeling combined with preset data to train the recognition model, significantly reducing labeling workload while maintaining effective model training.
Solution Approach 2:
The system merges preset data with on-site collected data to form a comprehensive training dataset. By combining complete labeling information from preset data with partial labeling from on-site data, the system achieves effective training without requiring complete manual labeling of all on-site objects.
3Measurement precision
If complete labeling information is required for all radiation images, then model training accuracy is improved, but labeling time and resources increase significantly
Solution Approach 1:
The system uses incomplete labeling information from on-site collected data combined with preset data. It accepts that not all objects need to be labeled and achieves effective model training through this partial labeling approach, significantly reducing the time and resources required for data preparation.
Solution Approach 2:
The system uses preset data as an intermediary to bridge the gap between incomplete on-site labeling and the need for comprehensive training data. The preset data provides the missing labeling information, enabling accurate model training without requiring complete manual labeling of all on-site images.
4Ease of operation
If the recognition system is deployed with fixed model, then initial recognition capability is provided, but adaptability to changing user needs is poor
Solution Approach 1:
The system transitions from a static deployed model to a dynamic self-updating system. The recognition model can be continuously trained and updated on-site using newly collected data combined with preset data, allowing the system to adapt dynamically to changing user needs and varying data distributions across different inspection sites.
Solution Approach 2:
The system implements a feedback loop where on-site collected data is used to retrain and update the recognition model. The system continuously learns from new data, evaluates model performance, and updates the model accordingly, enabling continuous improvement and adaptation to changing requirements.
Data Source
AI summary
Provided are a method and an apparatus of training a radiation image recognition model online and a method and an apparatus of recognizing a radiation image. Preset data is pre-stored in a radiation inspection system, and the preset data includes a first radiation image. The method of training radiation image recognition model online includes: collecting a second radiation image at the radiation inspection site; labeling the second radiation image online to form on-site collected data having a second labeling information, wherein the second labeling information is an incomplete labeling information of the second radiation image, and the second labeling information is used to indicate at least one of objects of interest in the second radiation image; synthesizing the preset data and the on-site collected data online to form a training data set; and training the radiation image recognition model online using the training data set.