An Augmentation Small Object Detection Method Based on NAS-FPN
Senhao Wang
Abstract
Aim at solving the problem that small object detection accuracy is poor, caused by fewer pixels contained and difficult feature extraction, and to improve pyramidal feature re-presentations, this paper proposes an effective method of augmented small detection based on a new feature pyramid architecture adopting Neural Architecture Search (NAS-FPN). We adopt scaling transformation, contrast enhancement, flipping, brightness alteration and rotating with a random angle as our augmentation approaches to process the input image data to obtain better small object detection effect. As a result, the accuracy of our methods tested on VOC2007 test set is relatively higher than Feature Pyramid Networks(FPN) and NAS-FPN, especially on small object detection such as potted plant, chair and dining table, as presented in TABLE.II.