Multi-Layer Image Classification System for Sub-Category Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image classification processes face challenges in accurately categorizing images, particularly when images belong to sub-categories, as existing methods often rely on a single layer of classifiers with uniform training processes, leading to inefficiencies in identifying specific sub-classes within broader categories.
Innovation Solution
The proposed solution involves a multi-layer image classification system with a first layer for initial categorization and a second layer for sub-category classification, using convolutional neural networks (CNNs) and Softmax classifiers to extract features and calculate probability values, allowing for independent training of classifiers within each layer to enhance recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single layer of classifiers with uniform training process is used, then the device complexity is reduced, but the classification accuracy for sub-categories deteriorates
Solution Approach 1:
The patent divides the image classification system into multiple layers: a first layer of classifiers for initial categorization and a second layer of classifiers for sub-category classification. This segmentation allows each layer to specialize in specific classification tasks, improving overall accuracy for sub-categories while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent introduces a hierarchical dimension to the classification system by organizing classifiers into multiple layers. The first layer handles broad categories while the second layer handles specific sub-categories, adding a dimensional structure that enables more precise classification without proportionally increasing overall system complexity.
2Measurement precision
If a multi-layer classification system with independent training processes is used, then the classification accuracy for sub-categories is improved, but the device complexity increases
Solution Approach 1:
The patent segments the classification system into distinct layers with independent training processes. Each layer can be trained separately on specific datasets, allowing for specialized optimization without requiring complex joint training mechanisms. This segmentation improves sub-category accuracy while managing complexity through independent module design.
Solution Approach 2:
The first layer of classifiers performs preliminary categorization before images are passed to the second layer for sub-category classification. This preliminary action filters and organizes images in advance, enabling the second layer to focus computational resources on specific sub-category differentiation, thereby improving accuracy without proportionally increasing overall complexity.
3Ease of manufacture
If uniform training process is applied to all classifiers, then the ease of manufacture is improved, but the recognition accuracy for closely related sub-categories deteriorates
Solution Approach 1:
The patent applies different training processes to different layers of classifiers based on their specific requirements. The first layer classifiers are trained on broad category data while the second layer classifiers are trained on specialized sub-category data. This local quality approach allows each layer to be optimized for its specific classification task, improving recognition accuracy for closely related sub-categories while maintaining ease of manufacture through standardized training procedures within each layer.
Data Source
AI summary
A method for classifying images by applying probabilities in steps includes selecting image category corresponding to an image classifier with the first probability value top ranked from the first layer image classifier set when the first layer image classifier set cannot classify an image to be identified. A second layer image classifier set is established according to the selected image category. The method selects output of the image classifier with the largest second probability value from the second layer image classifier set as result of classification when the second probability value of at least one image classifier is greater than or equal to a second predetermined value.