Data enhancement method for object detection. Issue 1 (1st February 2022)

• Record Type: Journal Article
• Authors: Yin, Zhijian; Xu, Lei; Yu, Hanqing; Yang, Zhen
• Is Part Of: Journal of physics. Volume 2209:Issue 1(2022)
• Publication Date: 2022-02-01
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/2209/1/012027
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: A data enhancement method for object detection is proposed to address the problem that there are few data enhancement means for target detection. This method has obvious advantages for small targets and scenarios where the target background is lacking. It is also a method to solve the problem of small and unbalanced samples. In this paper, the object is subtracted from the bounding box, and after processing, it is fused into a new image using various image classification improvement means such as random brightness, random contrast, random rotation, random cropping, etc. Using mixup-like fusion methods, and XIOU is proposed to guarantee the t intersection ratio of the target synthesis during synthesis. An improvement of 2% is obtained on yolov3 over the VOC dataset.