Image Classification Module for Tail-Class Feature Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing inspection systems struggle to accurately identify features in images, particularly in classes with insufficient sample images, known as tail-classes, leading to poor recognition and classification accuracy.
Innovation Solution
The system employs a classification module that generates supplemental images by combining class generic and specific features from head and tail classes, increasing the number of samples in tail-classes to improve training data variation and recognition accuracy, using a multi-class recognition process and deep convolutional neural networks for self-learning and image augmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the inspection system uses standard training methods with limited sample images, then the system can be trained quickly with available data, but the accuracy of identifying features in tail-classes is poor
Solution Approach 1:
The patent generates synthetic sample images by copying and combining features from existing images. The system extracts features from head-class images and reconstructs new images that resemble tail-class examples, effectively creating copies of useful feature patterns to augment scarce training data.
Solution Approach 2:
The patent introduces an intermediary feature space that mediates between head-class and tail-class images. By extracting and recombining features through this intermediate representation, the system bridges the gap between abundant head-class data and scarce tail-class data, enabling effective training despite data imbalance.
2Productivity
If the system focuses on learning from head-classes with sufficient samples, then the training process is efficient and converges faster, but the system fails to learn features of tail-classes accurately
Solution Approach 1:
The patent applies local quality by treating head-class and tail-class images differently in the training process. Head-class images are used to extract generic features that capture common patterns, while tail-class images are used to extract specific features that capture unique characteristics, allowing the system to adapt to each class's specific needs.
Solution Approach 2:
The patent segments the feature learning process into two distinct components: class-generic features learned from head-classes and class-specific features learned from tail-classes. This segmentation allows the system to efficiently learn from abundant head-class data while separately acquiring the specific knowledge needed for rare tail-classes.
3Device complexity
If the inspection system uses a simple classification approach, then the system complexity is low and easier to implement, but the system cannot accurately handle long-tail distribution of features
Solution Approach 1:
The patent changes the parameters of the classification approach by introducing separate feature extraction pathways for generic and specific features, and by using a two-stage training process. These parameter changes enable the system to handle long-tail distributions without requiring fundamentally complex architectural modifications.
Data Source
AI summary
A method of inspecting a component using an image inspection controller that includes a processor communicatively coupled to a memory includes classifying each sample image in a first database as a first sample or a second sample using a classification module, extracting at least one class generic feature from each first sample to generate a plurality of class generic features, and extracting at least one class specific feature from each second sample to generate a plurality of class specific features. The method further includes combining the class generic features and the class specific features to generate a plurality of supplemental images. The method further includes storing the sample images and the supplemental images in a second database, classifying each sample image and each supplemental image, capturing at least one image of the component using a camera, and identifying at least one feature of the component in the at least one image of the component using the classification module.