Function Approximator Normal Abnormal Classification Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems using function approximators, such as convolutional neural networks, lack the ability to distinguish between normal and abnormal classifications, making it difficult to calculate over detection and oversight rates, which are crucial for evaluating accuracy.
Innovation Solution
An information processing system that includes a reception unit for class registration, a model generation unit for function approximators, an estimation unit for image analysis, a determination unit for normal/abnormal classification, and an evaluation unit for calculating over detection and oversight rates, using response maps and threshold values to evaluate accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a function approximator is used for image recognition classification, then the recognition process can be performed, but it is impossible to distinguish between normal and abnormal classifications, making accuracy evaluation difficult
Solution Approach 1:
The patent segments the classification output into distinct normal and abnormal categories by introducing a determination unit that analyzes response map values. Instead of treating classification as a single undifferentiated output, the system divides it into separable normal/abnormal regions based on threshold comparisons, enabling subsequent accuracy evaluation.
Solution Approach 2:
The patent introduces response maps as an intermediary component between the function approximator and the final classification result. These response maps provide intermediate value information that serves as a bridge, allowing the determination unit to evaluate whether each region meets normal or abnormal criteria, thereby enabling accuracy measurement without modifying the core classification function.
2Device complexity
If traditional classification methods are used, then the image recognition process is simple, but over detection rate and oversight rate cannot be calculated
Solution Approach 1:
The patent performs preliminary action by generating response maps before final classification determination. These response maps are created in advance and store intermediate values that will be used later for accuracy evaluation, allowing the system to calculate over detection and oversight rates without requiring fundamental changes to the classification process.
Solution Approach 2:
The patent implements feedback by using the determination result (normal/abnormal classification) to evaluate the function approximator's performance. The evaluation unit compares determination results against actual ground truth data, providing feedback on over detection and oversight rates that can be used to assess and improve system reliability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A class registration unit 32 receives registration of a plurality of classes each corresponding to a normal attribute or an abnormal attribute. A model generation unit 36 generates a function approximator by learning based on a plurality of training images each corresponding to one of the plurality of classes. An estimation unit 38 uses the function approximator to estimate a portion, in each of a plurality of evaluated images registered as capturing a normal subject or an abnormal subject, corresponding to each class. A determination unit 40 determines whether a subject in each evaluated image is normal or abnormal based on a result of synthesizing portions in each evaluated image meeting a class corresponding to an abnormal attribute. An evaluation unit 42 evaluates the function approximator based on whether each evaluated image is registered as capturing a normal subject or capturing an abnormal subject and on a result of determining whether the subject in each evaluated image is normal or abnormal.