Image Identification Using Neural Network and Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image identification techniques struggle with accurately classifying input images as defective or non-defective when the difference in feature values between the input image and training images is either very large or very small, leading to reduced identification accuracy.
Innovation Solution
An image identification apparatus that combines a neural network for classification and an anomaly determination unit using a function approximator to calculate classification and anomaly scores, allowing for precise identification of input images as either similar to or dissimilar from training images, thereby improving accuracy across varying differences in feature values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classifier trained with labeled training images is used for image identification, then identification accuracy is improved when the difference in feature values between training images and input images is small, but identification accuracy deteriorates when the difference in feature values is remarkably large
Solution Approach 1:
The patent merges two different approaches: (1) a classifier trained with labeled training images that excels at identifying images with small feature value differences, and (2) an anomaly detection system that detects images with large feature value differences. By combining these two systems, the patent resolves the contradiction by allowing each system to handle the type of input it is best suited for, thereby achieving both high identification accuracy and adaptability to large variations.
Solution Approach 2:
The patent segments the image identification task into two distinct pathways: one for handling images with small feature value differences (using the trained classifier) and another for handling images with large feature value differences (using anomaly detection). This segmentation allows each subsystem to optimize for its specific function, resolving the contradiction between accuracy for similar images and adaptability for dissimilar images.
2Loss of information
If anomaly detection based on autoencoder is used to identify images with small feature value differences, then the degree of difference can be derived, but identification accuracy deteriorates when the difference between feature values is small making it difficult to identify defects
Solution Approach 1:
The patent combines anomaly detection (which derives the degree of difference) with a trained classifier (which provides high accuracy for defect identification). The system uses the anomaly detection to measure the degree of difference, then routes the input image to the appropriate processing pathway based on this measurement, thereby maintaining both information derivation capability and high defect identification accuracy.
Solution Approach 2:
The patent introduces an intermediary mechanism that measures the degree of difference between input images and training images. This intermediary assessment allows the system to determine which processing pathway to use, thereby enabling the anomaly detection system to provide useful information without compromising the high accuracy defect identification capability of the trained classifier.
3Productivity
If only classification based on trained neural network is used, then identification speed is maintained, but the system cannot accurately detect sudden defective products unprecedented in the past or images where the subject is not photographed
Solution Approach 1:
The patent performs preliminary action by pre-training the classifier with labeled training images to establish baseline identification capability and speed. This preliminary training allows the system to quickly process images that conform to known patterns, while the anomaly detection component is prepared to handle unprecedented cases. This preliminary preparation maintains identification speed for common cases while enabling detection of novel defects.
Solution Approach 2:
The patent introduces an intermediary anomaly detection system that acts as a supplement to the fast classification system. This intermediary component does not slow down the overall process significantly but provides the necessary adaptability to detect unprecedented defects and improperly photographed images, thereby resolving the contradiction between speed and versatility.
4Measurement precision
If dual-system approach combining classification and anomaly detection is used, then identification accuracy is improved across varying differences, but device complexity increases
Solution Approach 1:
The patent makes the system multi-functional by designing it to handle both classification-based identification and anomaly-based detection within a unified framework. The system can automatically select the appropriate function based on the input characteristics, thereby achieving high identification accuracy across varying differences without requiring completely separate systems, thus managing complexity while maintaining versatility.
Solution Approach 2:
The patent introduces dynamic selection between classification and anomaly detection pathways based on the characteristics of the input image. This dynamic approach allows the system to adapt its processing route in real-time, achieving high accuracy for different types of inputs while avoiding the complexity of permanently maintaining multiple parallel systems for all possible input types.
Data Source
AI summary
An image identification apparatus includes an acquisition unit configured to acquire an input image, a classification unit configured to calculate a classification score of the input image based on a neural network having a weight coefficient in each layer determined to calculate the classification score indicating a degree of similarity between a training image and an image to be processed, an anomaly determination unit configured to calculate an anomaly score of the input image based on a function approximator constructed by machine learning based on the training image of correct answer, and an identification unit configured to classify the input image into a good image having a high degree of similarity to the training image of the correct answer or a bad image having low degree of similarity to the training image of correct answer based on the classification score and the anomaly score.