Hierarchical Image Classification With Skip-Connected Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hierarchical image classification methods face accuracy issues as the number of classes increases and classification complexity grows.
Innovation Solution
A class prediction model utilizing a neural network architecture with skip connections between higher-level and lower-level classifiers, based on a VGG network, to enhance hierarchical classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If hierarchical classification is performed with increasing number of classes, then classification completeness is improved, but classification accuracy deteriorates
Solution Approach 1:
The classification task is segmented into multiple hierarchical levels (coarse-grained to fine-grained classes). Each level handles a specific number of classes, breaking down the complex multi-class problem into simpler sub-problems that can be solved with higher accuracy at each stage.
Solution Approach 2:
The patent introduces a hierarchical dimension to the classification structure, organizing classes in multiple levels from general to specific. This dimensional organization allows the system to manage increasing numbers of classes while maintaining accuracy by processing them through staged classification rather than a single flat classification step.
2Adaptability or versatility
If more classes are added to each layer, then classification coverage is improved, but classification accuracy deteriorates
Solution Approach 1:
The patent segments the large number of classes into multiple hierarchical layers, where each layer contains a manageable number of classes. This segmentation prevents any single classification layer from becoming overly complex, thereby maintaining accuracy while achieving comprehensive coverage through the combination of all layers.
Solution Approach 2:
The coarse-grained classification layers perform preliminary classification to narrow down the possible classes before the fine-grained layers conduct more detailed classification. This preliminary action reduces the search space for subsequent layers, improving overall accuracy while maintaining comprehensive coverage.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing apparatus includes an acquiring means for acquiring an object image that includes an object, and a predicting means for predicting a class of the object by applying the object image acquired by the acquiring means to a learning model. The learning model includes a main network (201) that extracts feature amounts regarding hierarchical classes of the object based on the object image, and a sub network (202) that outputs hierarchical classes of the object based on the feature amounts. The main network (201) includes extractors that extract the feature amounts. The sub network (202) includes classifiers that output classes of the object based on the feature amounts, and in the sub network, a higher-level classifier has a connection to one or more lower-level classifiers.