图像抠图与copy-paste结合的数据增强方法  

Data Enhancement Method Combining Image Matting and Copy-Paste

作　　者：杨天成 杨建红[1] 陈伟鑫 YANG Tiancheng;YANG Jianhong;CHEN Weixin(College of Mechanical Engineering and Automation,Huaqiao University,Xiamen 361021,China)

机构地区：[1]华侨大学机电及自动化学院,福建厦门361021

出　　处：《华侨大学学报（自然科学版）》2023年第2期243-249,共7页Journal of Huaqiao University(Natural Science)

基　　金：福建省科技重大专项(2020YZ017022);福建省厦门市科技计划项目(2021FCX012501190024);深圳市科技计划项目(JSGG20201103100601004)。

摘　　要：提出一种基于图像抠图与copy-paste结合的数据增强方法(matting-paste),采用图像抠图法获取单个垃圾实例的准确轮廓,并对单个实例进行旋转和亮度变换.根据物体轮廓信息,把实例粘贴到背景图上,无需额外的人工标注即可生成新的带有标注的数据,从而提高数据集的多样性和复杂性.结果表明:数据集扩充后的mask比数据集扩充前的识别精度提高了0.039,matting-paste能在已有数据集上有效地扩充数据,进一步提高模型的识别精度.A data enhancement method(matting-paste) based on image matting and copy-paste is proposed. Using the image matting method to obtain the precise contour of a single waste instance, and rotation and brightness transformation are carried out for each instance. Instances are pasted onto the background image according to the object′s contour information, and new annotated data can be generated without additional manual annotation, which improves the diversity and complexity of the dataset. The results show that the recognition precision of mask after dataset augmentation is improved 0.039 compared with before dataset augment. Matting-paste can effectively augment the data and further improve the the recognition precision of the model.

关 键 词：数据增强 图像抠图 COPY-PASTE 实例分割 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]