Classifier Generator Using Pre-trained Wide Residual Network Feature Extractor
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep neural network-based image classification methods require massive amounts of labeled data and struggle to adapt quickly when the image class changes, affecting classification accuracy.
Innovation Solution
A method involving pre-training a wide residual network as a feature extractor, randomly selecting classes and images, extracting feature vectors, and performing class information fusion and parameter prediction using a classifier generator to generate a classifier with a small number of labeled images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network-based classification methods are used, then classification accuracy is improved, but the requirement for massive amounts of labeled data increases
Solution Approach 1:
The patent applies preliminary action by pre-training a wide residual network on a large dataset to extract general image features, then freezing these features and only training the classifier head on the small number of labeled images. This allows the model to leverage pre-learned representations while adapting to the specific classification task with minimal data.
Solution Approach 2:
The patent segments the neural network into two functional parts: a feature extractor (pre-trained wide residual network) that handles general image understanding, and a classifier head that performs specific classification. This segmentation allows each part to be optimized independently, with the feature extractor trained on large data and the classifier trained on small data.
2Measurement precision
If deep neural network-based classification methods are used, then classification accuracy is improved, but the ability to adapt quickly to new classes deteriorates
Solution Approach 1:
The patent implements dynamics by making the classifier head trainable and adaptable while keeping the feature extractor frozen. This allows the system to quickly adapt to new classes by retraining only the lightweight classifier head on new labeled data, without retraining the entire heavy neural network.
Solution Approach 2:
The patent extracts the general feature extraction capability from the full neural network and separates it as a frozen component. This allows the system to maintain accurate feature representation while enabling rapid adaptation to new classification tasks through only the lightweight classifier head.
3Quantity of substance
If a classifier is trained on a small number of labeled images, then data requirement is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent introduces pre-trained feature vectors as an intermediary between the input images and the classifier. These feature vectors serve as a bridge that captures rich image information learned from large datasets, allowing the classifier to achieve high accuracy even when trained on only a small number of labeled images.
Data Source
AI summary
A method of generating a classifier by using a small number of labeled images includes: pre-training a wide residual network by using a set of labeled data with a data amount meeting requirements, and determining portions of the pre-trained wide residual network except for a fully connected layer as a feature extractor for an image; randomly selecting, for a N-class classifier to be generated, N classes from a training set for each of a plurality of times; and for N classes selected each time: randomly selecting one or more images from each class of the N classes as training samples; extracting a feature vector from training samples of each class by using the feature extractor; inputting a total of N feature vectors extracted into a classifier generator; and sequentially performing a class information fusion and a parameter prediction for the N-class classifier by using the classifier generator.