Image Object Attribute Classification Using Pseudo-Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning-based object/attribute classification systems require a large amount of labeled data for training, making them time-consuming and resource-intensive, and are limited in processing high-resolution images due to the constraints of pre-trained public convolutional neural networks.
Innovation Solution
An apparatus and method that utilize a combination of unsupervised and supervised learning to construct a classifier for image object attribute classification using a pre-trained public convolutional neural network, allowing for classification with a small amount of indexed data and enabling the processing of high-resolution images by dividing images into partial images for feature vector calculation and cluster formation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional deep learning-based classification systems are used, then classification accuracy can be achieved, but a large amount of indexed training data is required which increases time and effort for data preparation
Solution Approach 1:
The patent applies preliminary action by using unsupervised learning to pre-process un-indexed images and generate pseudo-labels before the main supervised classification task. This preliminary labeling step reduces the amount of manually indexed data needed while maintaining classification accuracy, directly addressing the time-consuming data preparation problem
Solution Approach 2:
The system performs self-service by automatically generating labels for training data through unsupervised learning algorithms. The model clusters un-indexed images and assigns cluster IDs as pseudo-labels without human intervention, enabling the system to create its own training data and reducing dependency on manually annotated datasets
2Measurement precision
If detailed and elaborated classification of features is performed, then classification precision is improved, but a larger amount of training data and expert index workers are required making data preparation more difficult
Solution Approach 1:
The patent performs preliminary clustering of un-indexed images using unsupervised learning to generate pseudo-labels for detailed attributes. This preliminary action creates structured training data that can be directly used for supervised fine-tuning, eliminating the need for expert annotators to manually create detailed attribute labels
Solution Approach 2:
The system automatically generates detailed attribute labels through unsupervised clustering algorithms that identify patterns in un-indexed images. This self-service approach replaces expert index workers with automated clustering, reducing both the complexity and cost of preparing detailed classification data
3Productivity
If pre-trained public convolutional neural networks are used, then training efficiency is improved, but they are limited in processing high-resolution images which are required for detailed attribute analysis
Solution Approach 1:
The patent segments high-resolution images into multiple patches or regions that can be processed by the pre-trained CNN architecture. This segmentation allows the model to handle high-resolution inputs by breaking them down into manageable chunks while preserving the ability to perform detailed attribute analysis on the original high-resolution data
Solution Approach 2:
The system transitions from directly processing high-resolution images in the spatial domain to processing image patches or feature representations in a different dimensional space. This dimensionality change allows the pre-trained network to effectively handle high-resolution inputs by operating on resized patches or extracted features rather than the full high-resolution tensor
4Quantity of substance
If a classifier is constructed with a small amount of indexed data, then resource requirements are reduced, but classification performance may deteriorate without sufficient training data
Solution Approach 1:
The system uses unsupervised learning to self-generate training data from un-indexed images by creating clusters and assigning pseudo-labels. This self-service data generation creates a synthetic training dataset that augmentsthe small amount of indexed data available, enabling the model to learn effective features without requiring large amounts of manually annotated data
Solution Approach 2:
The patent performs preliminary unsupervised learning on un-indexed images to generate pseudo-labeled training data before conducting supervised fine-tuning. This preliminary action creates a robust initial model that can be further refined with the small amount of indexed data, preventing performance deterioration that would occur with direct supervised learning on limited data
Data Source
AI summary
Provided is an apparatus for classifying an attribute of an image object, including: a first memory configured to store target object images that are indexed; a second memory configured to store target object images that are un-indexed; and an object attribute classification module configured to perform learning on the un-indexed target object images to construct a classifier for classifying a detailed attribute of target object, and finely adjust the classifier on the basis of the indexed target object images.