Image Classification via Partial Frame Labeling and Bounding Box Proposals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods for object localization and classification, such as convolutional neural networks, face challenges with high computational complexity and require extensive human-annotated data, which is laborious and insufficient for training, especially in accurately framing and classifying objects within images.
Innovation Solution
The proposed method involves training convolutional neural networks to label image crops as fully framed, partially framed, or background based on the presence of an object within a predetermined bounding box, using data augmentation techniques to increase training examples, and employing additional borders to improve object localization and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks are used for object localization and classification, then classification accuracy is improved, but computational complexity increases prohibitively
Solution Approach 1:
The patent segments the image processing task into multiple stages: first generating bounding box proposals, then classifying only those proposals. This divides the complex task of analyzing entire images into smaller, more manageable steps, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by generating bounding box proposals before classification. This pre-processing step identifies regions of interest in advance, so that subsequent classification only needs to process these limited regions rather than entire images, significantly reducing computational load.
2Measurement precision
If more labeled training data is collected for better model training, then classification accuracy is improved, but labor cost and time increase
Solution Approach 1:
The patent implements self-service by enabling the system to generate its own training data through automated bounding box proposal generation. The system uses unsupervised or weakly-supervised methods to create training examples without requiring extensive manual annotation, thus serving itself in data preparation.
Solution Approach 2:
The patent performs preliminary action by pre-processing images to generate bounding box proposals that can be used as training data. This preliminary processing creates structured training examples that are ready for classification model training, reducing the need for time-consuming manual data preparation.
3Measurement precision
If bounding box proposals are refined for better object framing, then localization accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies partial action by generating multiple bounding box proposals with varying degrees of precision. Rather than computing a single perfect bounding box, the system generates multiple proposals and selects the most appropriate ones for classification, achieving good localization accuracy with reduced computational effort.
Data Source
AI summary
A method of training for image classification includes labelling a crop from an image including an object of interest. The crop may be labelled with an indication of whether the object of interest is framed, partially framed or not present in the crop. The method may also include assigning a fully framed class to the labelled crop, including the object of interest, if the object of interest is framed. A labelled crop may be assigned a partially framed class if the object of interest is partially framed. A background class may be assigned to a labelled crop if the object of interest is not present in the crop.