Data Processing System Annotation Correction via Teacher Model Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing data processing system does not determine whether detected areas in inspection images include the target, leading to inefficiencies in label assignment and annotation correction.
Innovation Solution
A data processing system comprising a storage, an inferring section, an aiding section, and a correcting section, where the inferring section assigns inferred annotations to inspection images based on a teacher model, the aiding section assists humans in determining annotation correctness, and the correcting section generates corrected models by adding corrected annotations to the inferred annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the data processing system uses a teacher model to generate inferred annotations for inspection images, then the annotation process becomes automated and productivity increases, but the accuracy and reliability of the annotations deteriorate because the system does not perform determination based on the model as to whether or not each detected area includes the target
Solution Approach 1:
The system implements feedback by having the determination section evaluate the inferred annotations generated by the teacher model. The determination section uses the same teacher model to assess whether detected areas correctly include targets, providing a verification feedback loop that improves annotation reliability while maintaining automated productivity.
Solution Approach 2:
The determination section acts as an intermediary between the inference section and the final annotation output. It mediates by evaluating inferred annotations and determining their correctness before they are finalized, thus improving reliability without directly interfering with the automated annotation generation process.
2Reliability
If the system performs comprehensive determination for each detected area to ensure annotation accuracy, then reliability improves, but the complexity of the system and the time required for processing increase
Solution Approach 1:
The teacher model serves multiple functions: it generates inferred annotations and also performs determination/evaluation of those annotations. This multi-functionality allows the system to maintain reliability through comprehensive determination without adding separate specialized components, thus avoiding increased system complexity.
Solution Approach 2:
The system performs self-verification by using its own teacher model to evaluate the annotations generated by the same model. This self-service approach enables comprehensive determination without requiring external verification systems, maintaining simplicity while improving reliability.
3Measurement precision
If the system processes each inspection image through the teacher model for determination, then annotation accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The determination section does not need to process every single detected area with equal depth. It performs partial determination by evaluating whether detected areas include targets based on the teacher model's confidence and characteristics, achieving sufficient accuracy without exhaustive processing of all possible scenarios.
Solution Approach 2:
The teacher model generates inferred annotations with preliminary determination information before the formal determination process. This preliminary action includes initial assessments of detected areas, reducing the computational burden during the final determination phase and decreasing overall processing time.
Data Source
AI summary
Storage (24) stores therein a teacher model (54) that is a target image (50) with a teacher annotation (52), which specifies the target image (50), assigned thereto. An inferring section (32) assigns based on the teacher model (54) an inferred annotation (82) to an inspection image (80) belonging to the same category to which the target image (50) belongs. The inferred annotation (82) is a result of inference from the inspection image (80) based on the teacher model (54). An aiding section (34) aids a human to determine whether or not the inferred annotation (82) specifies the inspection image (80). A correcting section (36) generates a corrected model (86) in a manner to add a corrected annotation (84) to the inferred annotation (82) so that the inspection image (80) is specified.