Class-Aware Image Resolution Inference for Speed and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks struggle to maintain high prediction accuracy for optimal resolution conversion across various input images, as the resolution predictor fails to learn appropriate resolutions for each image effectively.
Innovation Solution
The proposed inference device and method cluster classes into multiple upper classes, determine a resolution for each class, predict the upper class, convert the image resolution accordingly, and perform classification, using a neural network-based resolution predictor and converter to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single resolution predictor is used to predict optimal resolution for all input images, then the device complexity is reduced, but the prediction accuracy deteriorates because it is difficult to learn the appropriate resolution for each image over all various images
Solution Approach 1:
The patent segments the prediction task by dividing multiple classes into multiple upper classes and creating separate resolution predictors for each upper class. This segmentation allows each predictor to specialize in predicting resolutions for specific types of images, thereby improving overall prediction accuracy while maintaining manageable system complexity through modular architecture
Solution Approach 2:
The patent introduces upper classes as intermediary categories between individual classes and resolution predictors. These upper classes serve as mediators that group related classes together, enabling the resolution predictors to learn patterns at an intermediate level of abstraction and improve prediction accuracy without requiring overly complex direct class-to-resolution mappings
2Productivity
If the resolution of input image is reduced to decrease computational load, then the processing speed is improved, but the inference accuracy deteriorates for samples that are easily blended with background
Solution Approach 1:
The patent implements dynamic resolution selection by using trained resolution predictors to determine the optimal resolution for each input image based on its characteristics. This dynamic approach allows the system to adaptively choose between lower resolutions (for speed) and higher resolutions (for accuracy) depending on the specific image content, thereby resolving the contradiction between inference speed and accuracy
Solution Approach 2:
The patent changes the resolution parameter adaptively based on the predicted upper class of the input image. By training multiple resolution predictors with different resolution settings and selecting the appropriate predictor based on image characteristics, the system optimizes the resolution parameter to balance computational load and inference accuracy for different types of images
Data Source
AI summary
The inference device converting the resolution of an input image and performing inference, includes a clustering unit which clusters multiple classes to be classified into multiple upper classes, a resolution determination unit which determines a resolution corresponding to each of the multiple upper classes, a prediction unit which predicts the upper class to which the class to be classified in the input image belongs, a resolution converter which converts the resolution of the input image to a resolution corresponding to the predicted upper class, and a classifier which performs classification on the input image whose resolution has been converted.