Image Classification Confidence Gating With Converted-Image Ensemble
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recent deep learning models trained on large-scale data exhibit black box characteristics, leading to insufficient accuracy when encountering images with characteristics different from those in the training dataset, and the causal relationship between input images and inference results is not fully elucidated.
Innovation Solution
An image classification apparatus that utilizes an image classification model, a confidence level evaluation unit, an image conversion unit, and an integrated evaluation unit to generate converted images when confidence is low, combining prediction vectors from the input and converted images to improve classification accuracy through ensemble learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single deep learning model is used for image classification, then the device complexity is reduced, but the reliability of classification decreases when encountering images with characteristics different from training data
Solution Approach 1:
The patent combines multiple inference engines (deep learning models) into an ensemble system where their prediction vectors are aggregated. This merging of multiple models improves classification reliability by reducing the impact of individual model limitations while maintaining manageable complexity through systematic integration.
Solution Approach 2:
The patent segments the classification task by dividing it into multiple independent inference engines that each process the input image separately. Each model produces its own prediction vector, which are then combined to form a final classification decision, allowing diversity in model architectures and training data while improving overall reliability.
2Measurement precision
If multiple inference engines are used for ensemble learning, then the classification accuracy improves, but the processing time increases
Solution Approach 1:
The patent performs preliminary actions by pre-training multiple inference engines on different datasets or with different architectures before deployment. During actual classification, these pre-prepared models can process images in parallel, reducing the time penalty of ensemble learning while maintaining the accuracy benefits of multiple perspectives.
Solution Approach 2:
The patent implements a confidence-based mechanism where if the initial classification confidence is sufficient, additional inference engines may be skipped. This partial action approach maintains high accuracy when needed while reducing processing time when the first model already provides a confident correct classification.
3Adaptability or versatility
If data augmentation is applied to expand training data, then the model's adaptability improves, but the training complexity increases
Solution Approach 1:
The patent creates multiple copies of the training dataset through data augmentation techniques, generating transformed versions of original images with different transformations (rotation, flipping, cropping, etc.). This copying approach expands the effective training data size and improves model adaptability without requiring collection of new real-world data, managing training complexity through systematic transformations.
Data Source
AI summary
In an image classification apparatus adapted to classify an input image, an image classification model calculates a prediction vector predicting a class into which the input image is classified. A confidence level evaluation unit calculates a probability vector and a confidence level of classification, based on the prediction vector for the input image. An image conversion unit converts the input image to generate a converted image when the confidence level is less than a first threshold value. An integrated evaluation unit calculates an integrated prediction vector derived from adding up the prediction vector for the input image and a prediction vector for the converted image calculated based on the image classification model and calculates an integrated probability vector from the integrated prediction vector. A classification determination unit determines a class of the input image based on the probability vector or the integrated probability vector.