一种基于生成式对抗网络的数据集扩充方法在湿地鸟类图像识别中的应用  

Application of a Data Set Expansion Method Based on Generative Adversarial Network in Wetland Bird Image Recognition

作　　者：熊聪聪[1] 赵子健 曹鉴华[1] 汤凌燕 XIONG Congcong;ZHAO Zijian;CAO Jianhua;TANG Lingyan(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2022年第4期64-70,共7页Journal of Tianjin University of Science & Technology

基　　金：天津市教委基金项目(2018KJ-106)。

摘　　要：已有的天津大黄堡湿地鸟类数据集存在样本数量少、样本分布不均衡等问题,严重影响鸟类智能识别的精度.本文提出一种基于生成式对抗网络的数据集扩充方法,通过使用旋转、缩放、添加噪声等方式进行初步扩充,然后输入条件对抗网络进行学习,通过使用生成器生成新的鸟类样本.对大黄堡湿地鸟类数据集进行扩充处理,并使用目标检测算法SSD、YOLOv3-SPP-Net、RetinaNet进行测试,结果表明识别的精度显著提高,基本满足本区鸟类目标识别的需求.The existing bird data set of Dahuangbao wetland in Tianjin has such problems as small sample number and uneven sample distribution,which have seriously affected the accuracy of bird intelligent recognition.In this study we propose a sample expansion method based on affine transformation and adversarial network,namely,preliminarily expanding by using rotation,scaling and noise,and then inputting the conditional adversarial network for learning,and using the generator to generate new bird pictures.The bird data set of Dahuangbao wetland was expanded and tested with three target detection algorithms:SSD,YOLOv3-SPP-Net and RetinaNet.The test results showed that the recognition accuracy significantly improved,which basically met the needs of bird target recognition in this area.

关 键 词：鸟类识别 数据扩充 对抗网络 目标检测 

分 类 号：TP3-05[自动化与计算机技术—计算机科学与技术]