研究目的
To address the challenge of annotating sufficient amount of data for training object detection networks by proposing a semi-supervised learning based method that utilizes Generative Adversarial Networks to extract data distribution from unannotated data.
研究成果
The semi-supervised learning approach using GAN effectively reduces false alarms and makes good utilization of data, while preventing overfitting. However, it does not address the issue of missing detection rates.
研究不足
The approach cannot reduce the missing detection rate, requiring the detector to ensure a low missing detection rate.
1：Experimental Design and Method Selection:
The study employs a semi-supervised learning approach using Generative Adversarial Networks (GAN) to improve object detection performance with few annotated data and massive unannotated data.
2：Sample Selection and Data Sources:
Over 1000 airplanes from images of 11 airports in China and America are collected, with one airport used as testing dataset, 100 airplanes as labeled dataset, and the rest as unlabeled dataset.
3：List of Experimental Equipment and Materials:
The study uses DRBox as the detection network and a GAN based on DCGAN for semi-supervised learning.
4：Experimental Procedures and Operational Workflow:
The detection network is first trained with labeled samples, then used to detect objects in unannotated images. The detection results are used as unlabeled data in the training of the semi-supervised classification network.
5：Data Analysis Methods:
The performance is evaluated using precision-recall (P-R) curves and metrics such as recall ratio and false alarms ratio.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
