Recurrent Neural Network for Video Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The process of labeling training images for neural networks is time-consuming, costly, and laborious, and not all images are available for public dissemination due to legal or contractual restrictions, making it difficult to analyze images effectively, especially in applications like medical imaging or object recognition in restricted domains.

Innovation Solution

An image analysis system and method that combines information from multiple image frames to improve neural network comprehension, using recurrent neural networks like LSTM to provide temporal context for accurate object prediction and tracking without additional labeled training images, allowing for rapid and accurate identification of objects in video data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If labeled training images are used to train neural networks, then the accuracy of object detection is improved, but the time and cost required for labeling increases significantly

Engineering Contradiction:
Improveobject detection accuracyVSAvoidlabeling time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by using semi-supervised learning to pre-train neural networks with limited labeled data and unlabeled data, preparing the model in advance so that less time is needed during actual object detection tasks. The neural network is pre-adapted to the specific domain and data distribution before deployment.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces semi-supervised learning as an intermediary approach between fully supervised and unsupervised learning. This intermediary method uses a small amount of labeled data combined with large amounts of unlabeled data, mediated through techniques like pseudo-labeling and consistency regularization, to achieve good detection accuracy without requiring extensive manual labeling.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If more labeled training images are collected, then the neural network's ability to detect various objects improves, but the cost and labor requirements increase

Engineering Contradiction:
Improveobject detection capabilityVSAvoiddata collection cost
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The neural network performs self-service by generating its own training data through pseudo-labeling. The model predicts labels for unlabeled data, and these predictions are used as pseudo-labels to further train the network. This self-generated labeling process eliminates the need for expensive manual annotation while expanding the network's object detection capabilities.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent merges labeled and unlabeled data into a unified training framework. By combining a small set of labeled images with a large set of unlabeled images in semi-supervised learning, the system achieves versatile object detection capability without the high costs associated with labeling large datasets. The merging is facilitated by techniques like consistent pseudo-labeling and multi-criteria filtering.

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If conventional neural networks are used for image analysis, then the system complexity is low, but the ability to handle temporal relationships and restricted data is insufficient

Engineering Contradiction:
Improvesystem complexityVSAvoidtemporal relationship handling
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent adds the temporal dimension to image analysis by processing sequences of images or video frames rather than single static images. This dimensional extension allows the neural network to capture temporal relationships and motion patterns, improving reliability in applications like action recognition and video analysis without significantly increasing system complexity.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system transitions from static image analysis to dynamic video analysis by implementing temporal modeling in neural networks. This allows the system to adapt to changing scenes and objects over time, improving reliability for temporal relationship handling while maintaining manageable complexity through efficient temporal processing architectures.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3407253B1Image analysis neural network systems
Publication Date: 2024.07.03 GENERAL ELECTRIC CO
  • EP3407253B1 patent drawingFigure 1
  • EP3407253B1 patent drawingFigure 2~3
  • EP3407253B1 patent drawingFigure 4

AI summary

A method includes obtaining a series of images (503) of a rotating target object (210, 310) through multiple revolutions of the target object (210, 310). The method includes grouping the images (503) into multiple, different sets (502, 504) of images (503). The images (503) in each of the different sets (502, 504) depict a common portion of the target object (210, 310). At least some of the images (503) in each set (502, 504) are obtained during a different revolution of the target object (210, 310). The method further includes examining the images (503) in at least a first set (502) of the multiple sets (502, 504) of images (503) using an artificial neural network (102) for automated object-of-interest recognition by the artificial neural network (102).