Neural Network Path Extraction for Image Classification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks face challenges in accurately classifying images, particularly in distinguishing between conforming and defective articles, leading to high 'overlooking' rates and operational difficulties in adapting the network for such classifications.
Innovation Solution
An image classification system that extracts processing paths in a network by connecting feature values and fully connected output values, including maximum values, to compare with pre-set paths and determine the associated class for input images, enhancing classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural network is used for image classification, then identification rate is improved, but overlooking rate (erroneous determination of defective articles as conforming) increases
Solution Approach 1:
The patent segments the neural network processing into distinct path extraction and path comparison stages. By extracting specific processing paths that connect feature values to output values and comparing them against pre-stored paths, the system divides the classification task into manageable segments that reduce overlooking errors while maintaining identification rate.
Solution Approach 2:
The system implements feedback by storing multiple processing paths associated with different classes and using the comparison results to verify classifications. When a mismatch is detected between extracted and stored paths, the system can identify potential overlooking errors and adjust accordingly, creating a feedback loop that improves reliability.
2Adaptability or versatility
If convolutional neural network is adapted for determining conforming or defective articles, then classification capability is improved, but operational complexity increases due to network reconstruction requirements
Solution Approach 1:
The patent applies preliminary action by pre-storing multiple processing paths in the storage unit before actual classification operations. These pre-stored paths represent different class characteristics and are prepared in advance, eliminating the need for network reconstruction when new classes need to be evaluated, thus reducing operational complexity while maintaining adaptability.
Solution Approach 2:
The system creates copies of processing paths for different classes and stores them in the storage unit. Instead of reconstructing the neural network for each classification task, the system uses stored path copies that can be quickly retrieved and compared, significantly reducing operational complexity while preserving classification capability.
3Measurement precision
If maximum value selection is used in path extraction, then classification accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent applies local quality by selecting maximum values at specific critical points in the processing path (feature values and output values) rather than processing all values uniformly. This localized maximum value selection improves classification accuracy by focusing on the most significant features while keeping processing complexity manageable through selective rather than comprehensive processing.
Data Source
AI summary
Provided is an image classification system for classifying an input image including: a network constructed so that feature amount extraction processing and fully connected processing are executed on input image data; a path extracting means configured to extract a processing path in the network; and a determination module configured to compare the processing path extracted by the path extracting means and a plurality of setting paths set in advance in association with classes to set a class associated with a matched path as a class of the input image.