Object Detection Network Learning with Supervised and Self-Supervised Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolutional neural networks (CNNs) require abundant data for excellent performance, making it difficult to improve detection accuracy when using a small amount of learning data.
Innovation Solution
A learning device with a first learning unit that includes a supervised learning unit and a self-supervised learning unit, which learns an object detection network using both supervised and self-supervised data to reduce losses in object detection and feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If abundant supervised data is used for training object detection networks, then detection accuracy is improved, but data collection cost and time consumption increase
Solution Approach 1:
The patent introduces self-supervised learning as an intermediary mechanism that generates pseudo-labels from image data without requiring manual annotation. This intermediary process allows the system to create training data from existing images, thereby reducing the need for abundant manually labeled supervised data while maintaining detection accuracy.
Solution Approach 2:
The system performs self-supervised learning where the model generates its own training labels by processing image data through pre-trained features. This self-service mechanism enables the system to create supervisory signals from unlabelled data, eliminating the dependency on abundant manually annotated data while preserving detection performance.
2Quantity of substance
If self-supervised learning is applied to generate training data, then amount of learning data is reduced, but detection accuracy for new classes deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-training the object detection network on abundant data from existing classes before fine-tuning on new classes with limited data. This pre-training establishes robust feature representations and detection capabilities that can be leveraged during subsequent fine-tuning, enabling accurate detection of new classes even with small amounts of labelled data.
Solution Approach 2:
The system dynamically adjusts learning parameters and loss weights during training to prioritize features relevant to new classes while maintaining general detection capabilities. By changing parameter configurations based on the specific task requirements and available data, the model adapts to new classes effectively without requiring abundant training data.
Data Source
AI summary
A learning device 10 includes a first learning unit 20. The first learning unit 20 includes a first supervised learning unit 22 and a first self-supervised learning unit 24. The first supervised learning unit 22 learns a first object detection network 30 using learning data 40 so as to reduce a first loss between an output of the first object detection network 30 for detecting an object from target image data and supervised data 40B. Using image data 40A and self-supervised data 40C generated from the image data 40A, the first self-supervised learning unit 24 learns the first object detection network 30 so as to reduce a second loss of a feature amount of a corresponding candidate area P between the image data 40A and the self-supervised data 40C, the second loss being derived by the first object detection network 30.