基于自适配归一化的改进Mask Scoring R-CNN  

Improved Mask Scoring R-CNN based on Switchable Normalization

作　　者：张幸 赵文仓[1] 王旭 Zhang Xing;Zhao Wencang;Wang Xu(College of Automation and Electronics Engineering,Qingdao University of science&Technology,Qingdao 266061,China)

机构地区：[1]青岛科技大学自动化与电子工程学院,青岛266061

出　　处：《电子测量技术》2020年第6期93-98,共6页Electronic Measurement Technology

基　　金：国家留学基金委项目(201608370049);国家自然科学基金(61171131);山东省重点研发计划(YD01033)项目资助。

摘　　要：基于批量归一化的mask scoring R-CNN在目标检测与实例分割领域展现出卓越性能,其平均精度明显高于传统实例分割模型Mask R-CNN。但是由于批量归一化方法存在小批量精度骤降和大批量GPU内存溢出的缺陷,影响到实际应用中的检测与分割任务效果。自适配归一化方法对各批量大小都有极佳的鲁棒性,可以弥补上述不足。从数学角度给出了减少自适配归一化中计算冗余的证明,并将其应用于mask scoring R-CNN,小批量条件下在COCO数据集内将检测精度提升了4.4%,分割精度提升了3.9%,进一步提升了模型性能。Mask scoring R-CNN based on batch normalization shows excellent performance in object detection and instance segmentation tasks,and its average precision was significantly higher than that of mask R-CNN,a traditional benchmark of instance segmentation framework.However,due to the shortcomings of batch normalization,such as the precision decrease sharply in small batch size and out of memory in GPU in large batch size,that affects the performance of detection and segmentation performance in practical applications.Switchable Normalization,which has excellent robustness to all batch sizes.This work gives a proof to reduce the computational redundancy in switchable normalization from the mathematical point of view,and apply it to mask scoring R-CNN.Under the condition of small batch size,the detection average precision is improved by 4.4%and the segmentation average precision is improved by 3.9%in COCO,which further improves the model performance.

关 键 词：自适配归一化 批量归一化 目标检测 实例分割 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]