Hierarchical Neural Network for Image Classification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image classification methods in computer vision struggle to achieve accurate classification results, especially when dealing with images that belong to classes with limited training data or when the classification requires consideration of class hierarchies.
Innovation Solution
The proposed method involves an image classification approach that utilizes a neural network model trained with both label information and class hierarchy information of training images. This method classifies images by determining their class or superclass based on the trained neural network model, which considers the hierarchical relationships between classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image classification methods are used, then the classification process is simple, but the classification accuracy is insufficient especially for classes with limited training data
Solution Approach 1:
The patent implements a hierarchical classification structure where a superclassification network and subclassification networks are nested. The superclassification network first identifies the superclass of an image, then the corresponding subclassification network performs fine-grained classification. This nested architecture allows the system to leverage shared features across multiple classes while maintaining specialized classification capabilities, thereby improving accuracy for classes with limited training data without requiring each class to have an independently complex model.
Solution Approach 2:
The superclassification network serves multiple functions: it classifies images into superclasses and provides shared feature representations that are reused by multiple subclassification networks. This multi-functional design improves efficiency and accuracy by allowing the system to learn generalizable features at the superclass level that benefit all subordinate classes, reducing the data hunger of individual class classifiers.
2Measurement precision
If class hierarchy information is incorporated into training, then the classification accuracy improves, but the data processing complexity increases
Solution Approach 1:
The patent segments the classification task into two distinct stages: superclassification and subclassification. Class hierarchy information is processed differently at each stage - the superclassification network processes images to determine superclass labels, while subclassification networks process images within specific superclass contexts. This segmentation allows the system to effectively utilize hierarchical relationships without overwhelming the processing system with all hierarchy information simultaneously, managing complexity through staged processing.
Solution Approach 2:
The system applies different processing qualities to different parts of the hierarchy. The superclassification network uses broader, more general features suitable for high-level classification, while subclassification networks use more specialized, detailed features appropriate for fine-grained discrimination. This local quality adaptation allows effective use of hierarchical information at each level without requiring uniform complex processing throughout the entire hierarchy.
3Measurement precision
If more training data is collected for each class, then the classification accuracy improves, but the data collection time and cost increase
Solution Approach 1:
The patent merges multiple subclassification networks under a common superclassification framework. By combining the training process across related classes through the shared superclass structure, the system can leverage data from multiple classes to improve the classification of individual classes. This merging approach reduces the effective data requirement per class by allowing knowledge transfer and shared feature learning across the hierarchy, thereby maintaining accuracy while reducing data collection time and costs.
Data Source
AI summary
An image classification method, a neural network training method, and an apparatus are provided, and relate to the field of artificial intelligence, and specifically, to the field of computer vision. The image classification method includes: obtaining a to-be-processed image; and obtaining a classification result of the to-be-processed image based on a pre-trained neural network model, where the classification result includes a class or a superclass to which the to-be-processed image belongs. When the neural network model is trained, not only labels of a plurality of training images but also class hierarchy information of the plurality of training images is used. That is, more abundant information of the training images is used. Therefore, images can be better classified.