Semi-supervised Object Detection via Segmented Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems using semi-supervised learning face over-training issues, leading to a decline in generalization performance due to the combination of supervised and unsupervised data learning, particularly in feature extraction and object detection neural networks.

Innovation Solution

The object detection device employs a semi-supervised learning approach where the feature extraction neural network learns using unsupervised data and supervised data separately, and the object detection neural network learns using supervised data, while the original image restoration neural network learns using unsupervised data, thereby preventing over-training by associating object detection results with image regions and using meta-information for image restoration.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If semi-supervised learning is applied to both feature extraction and object detection neural networks using both supervised and unsupervised data, then the use of data is maximized, but over-training occurs and generalization performance declines

Engineering Contradiction:
Improveamount of training data usedVSAvoidgeneralization performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The learning system is segmented into two distinct neural networks with different learning strategies: an autoencoder for feature extraction that uses unsupervised learning on unlabeled data, and an object detection network that uses supervised learning on labeled data. This segmentation prevents over-training by applying appropriate learning methods to each component, thereby maintaining generalization performance while utilizing both supervised and unsupervised data effectively.

Inventive Principle:
Principle #1Segmentation

2Reliability

If only supervised data is used for object detection neural network learning, then over-training is suppressed, but the amount of useful data is limited

Engineering Contradiction:
Improvegeneralization performanceVSAvoidamount of training data
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The autoencoder serves as an intermediary that processes unlabeled data and generates meaningful feature representations. These features are then fed into the supervised object detection network. This intermediary approach allows the system to utilize large amounts of unlabeled data while maintaining the supervised learning framework that ensures generalization performance.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If supervised data is used for feature extraction learning, then detection accuracy is improved, but manpower cost increases due to data annotation requirements

Engineering Contradiction:
Improvedetection accuracyVSAvoidmanpower cost
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

Different quality levels of data are applied to different parts of the system: unlabeled data is used for the autoencoder's feature extraction component where perfect annotations are not necessary, while labeled data is reserved for the object detection network where precision is critical. This local quality approach reduces annotation costs while maintaining detection accuracy.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3410351B1Learning program, learning method, and object detection device
Publication Date: 2023.12.27 FUJITSU LTD
  • EP3410351B1 patent drawingFigure 1
  • EP3410351B1 patent drawingFigure 2
  • EP3410351B1 patent drawingFigure 3

AI summary

An object detection device (10) extracts feature for input data utilizing an encoder, the input data including labeled data and unlabeled data and detects object in each of the input data, utilizing an object detector. The object detection device (10) generates region data for each of the input data, each of the region data corresponding to the detected object and generates restoration data from the region data and meta-information related to the detected object for each of the input data utilizing a decoder corresponding to the encoder. The object detection device (10) executes learning of the encoder and the object detector based on a result detected by the object detector and a label associated with the input data, when the input data is labeled data, and executes learning of the encoder, the object detector, and the decoder, based on the input data and the restoration data.