Weakly Labeled Image Concept Discovery via Iterative Hard Instance Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Building large labeled datasets for computer vision recognition systems is costly and inaccurate due to the need for manual labeling or the inclusion of outlier images from automated processes like image search engines, which result in lower accuracy and require extensive training.
Innovation Solution
The use of weakly labeled image collections, where tags or descriptions from photo sharing websites are treated as labels, and an iterative hard instance learning algorithm to identify visual concepts, allowing for the creation of a training set for classifying images and providing image classification services.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to build large labeled datasets, then labeling accuracy is improved, but cost and time consumption increase
Solution Approach 1:
The patent introduces an automated preprocessing system that acts as an intermediary between raw image data and manual labeling. This system automatically filters and pre-processes images before they reach human labelers, reducing the volume of images requiring manual attention while maintaining high accuracy standards.
Solution Approach 2:
The patent implements preliminary automated filtering and organization of images before manual labeling begins. By pre-processing the dataset to identify and remove obvious outliers or low-quality images in advance, the system reduces the time required for manual labeling while preserving accuracy.
2Ease of manufacture
If image search engines are used to build labeled datasets, then cost is reduced, but accuracy deteriorates due to outlier images
Solution Approach 1:
The patent converts the harmful effect of noisy, outlier-containing data from automated search engines into a benefit by implementing an automated outlier detection and filtering system. This system identifies and removes problematic images while retaining the cost advantages of automated data collection, ultimately improving dataset quality.
Solution Approach 2:
The patent replaces manual inspection and filtering of automated search results with an automated computer vision-based filtering system. This substitution maintains low costs while significantly improving accuracy by systematically identifying and removing outlier images that would otherwise contaminate the dataset.
3Productivity
If automated image search is used to collect images, then productivity increases, but measurement precision decreases due to included outlier images
Solution Approach 1:
The patent implements a continuous automated filtering process that operates throughout the data collection pipeline. Rather than performing filtering as a separate post-processing step, the system continuously identifies and removes outliers as images are being collected, maintaining both high productivity and image quality throughout the entire process.
Solution Approach 2:
The patent substitutes manual quality review processes with automated computer vision-based filtering that operates at scale. This replacement maintains the high productivity of automated collection while systematically ensuring image quality through algorithmic outlier detection and removal.
Data Source
AI summary
Images uploaded to photo sharing websites often include some tags or sentence descriptions. In an example embodiment, these tags or descriptions, which might be relevant to the image contents, become the weak labels of these images. The weak labels can be used to identify concepts for the images using an iterative hard instance learning algorithm to discover visual concepts from the label and visual feature representations in the weakly labeled images. The visual concept detectors can be directly applied to concept recognition and detection.