[2D4-GS-2] Machine learning: Image recognition

[2D4-GS-2-05] Improving the Performance of Object Detection Using Self-supervised Learning

〇Taisuke Usumi¹, Shunsuke Kurumatani¹, Takeharu Eda¹ (1. NTT Software Innovation Center)

Keywords:Self-supervised Learning, Object Detection, Deep Learning

In recent years, as a market for image recognition AI has expanded, creating highly accurate AI models has become more important. However, since a large amount of labeled data is necessary to create highly accurate AI models in general, labeling costs become a significant issue. Hence, self-supervised learning which is able to train feature extractors using only unlabeled images has attracted a lot of attention in recent years. Therefore, we aim to create highly accurate object detection models that are commonly used in connected cars and smart cities using self-supervised learning. However, current self-supervised learning methods cannot train well when training with complex images where distinguishing between objects and backgrounds is difficult. We believe that this is due to a low accuracy of unsupervised object detection used in self-supervised learning for such complex images. Therefore, we propose a method to enable a feedback of unsupervised object detection to be used for training model. As a result of experiments, we improve the accuracy of unsupervised object detection, showing the possibility to create a highly accurate object detection models.