Convolutional Neural Network Training with Multi-Granularity Annotations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial intelligence techniques for classifying and analyzing digital images are inefficient and labor-intensive, requiring pixel annotations and voxel level annotations, which limits their accuracy and efficiency.
Innovation Solution
A system that includes a training component for a convolutional neural network using masks, image-level labels, and bounding boxes to generate loss functions, which are iteratively back-propagated to tune the network's parameters, improving classification and localization accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional artificial intelligence techniques are used for image classification and analysis, then the system can process digital images, but the accuracy and efficiency are difficult to achieve and require labor-intensive pixel annotations and voxel level annotations
Solution Approach 1:
The patent segments the annotation process by introducing multiple types of annotations (pixel annotations, voxel annotations, bounding box annotations, image-level labels) that can be used in different combinations. This segmentation allows the system to achieve high classification accuracy through pixel-level precision while improving productivity by optionally using coarser annotations like bounding boxes or image-level labels when full pixel annotation is not necessary.
2Measurement precision
If pixel annotations and voxel level annotations are used to improve classification accuracy, then measurement precision is improved, but the process becomes labor-intensive and time-consuming
Solution Approach 1:
The patent applies partial action by allowing the system to use only the necessary level of annotation detail for each specific classification task. Instead of always requiring complete pixel-level annotations, the system can use partial annotations such as bounding boxes or image-level labels when sufficient, reducing annotation time while maintaining adequate classification precision. The loss function is designed to work with whatever annotation level is provided.
Solution Approach 2:
The patent changes the parameter of annotation granularity by supporting multiple annotation types at different levels of detail. The system can dynamically adjust which annotation parameters are used based on the specific application requirements, allowing faster annotation processes for tasks that don't require maximum precision while still providing the option for full pixel-level annotation when needed.
3Measurement precision
If multiple types of annotations (masks, bounding boxes, image-level labels) are utilized with multiple loss functions, then classification and localization accuracy is improved, but the system complexity increases
Solution Approach 1:
The patent merges multiple annotation types and loss functions into a unified training framework. Instead of treating pixel annotations, voxel annotations, bounding box annotations, and image-level labels as separate systems, the patent combines them into a single neural network training process that uses a composite loss function. This merging reduces system complexity by providing a unified interface while maintaining the ability to achieve high localization accuracy through the coordinated use of all annotation types.
Solution Approach 2:
The patent creates a universal annotation system where a single neural network architecture and training framework can handle multiple types of annotations (pixel, voxel, bounding box, image-level labels) and multiple loss functions simultaneously. This multi-functional approach simplifies the system by eliminating the need for separate processing pipelines for different annotation types, while still achieving high classification and localization accuracy through the flexible use of whatever annotation types are available.
Data Source
AI summary
Systems and techniques for classification based on annotation information are presented. In one example, a system trains a convolutional neural network based on training data and a plurality of images. The plurality of images is associated with a plurality of masks, a plurality of image level labels, and/or a bounding box. The system also generates a first loss function based on the plurality of masks, a second loss function based on the plurality of image level labels, and a third loss function based on the bounding box. Furthermore, the system generates a fourth loss function based on the first loss function, the second loss function and the third loss function, where the fourth loss function is iteratively back propagated to tune parameters of the convolutional neural network. The system also predicts a classification label for an input image based on the convolutional neural network.