引入抑制增长损失函数的肺炎目标检测  

Introducing inhibition growth loss function to pneumonia object detection

作　　者：滕皓 陆慧娟 朱海天 朱文杰 TENG Hao;LU Hui-juan;ZHU Hai-tian;ZHU Wen-jie(College of Information Engineering,China Jiliang University,Hangzhou 310018,China;College of Economics and Management,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]中国计量大学信息工程学院,浙江杭州310018 [2]中国计量大学经济管理学院,浙江杭州310018

出　　处：《计算机工程与设计》2021年第9期2664-2670,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61272315、61602431);浙江省自然科学基金项目(LQ20F030015);中国计量大学卓越学生成果培育计划基金项目(2019YW16)。

摘　　要：为提高对肺炎图像目标检测算法的平均精度,针对肺炎图像中病灶的轮廓模糊、通道单一的缺点,提出抑制增长损失函数(inhibition growth loss function,IG Loss)。根据损失函数数值的变化特征自适应调整其权值,使损失函数值随step增大的项得到增强,损失函数值随step减小的项不变,引入激活函数Mish代替ReLU,减少信息的丢失。实验结果表明,在没有增加时间开销的情况下,基于该抑制增长损失函数的Faster R-CNN和R-FCN算法的AP分别提升3.8%和2.8%,验证了该抑制增长损失函数的有效性。To improve the average precision of the object detection algorithm for pneumonia images,in view of the shortcomings of the contour of the object in the pneumonia image and the single channel,an inhibition growth loss function was proposed.The weight of the loss function was adjusted adaptively according to its changing characteristics,so that the item whose value of loss function increased with step was enhanced,and the item whose value of loss function decreased with step was unchanged,and the activation function Mish was introduced instead of ReLU to reduce the loss of information.Experimental results show that the APs of the Faster R-CNN and R-FCN algorithms based on the inhibition growth loss function are increased by 3.8%and 2.8%,respectively,without increasing time overhead,and the effectiveness of the inhibition growth loss function is verified.

关 键 词：损失函数 目标检测 卷积神经网络 肺炎图像 自适应权值 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]