摘要
在基于深度学习的目标检测领域,借助图像特征提取技术的进步,可以使用锚框在图像的不同位置生成边框,通过提取边框区域特征进行边框位置回归。在实际训练过程中,发现基于anchor的单阶段检测方法存在精度不足的问题,经过分析总结将问题划分为3类:样本分类不均衡;目标检测多尺度;优化目标与推论不一致。经过研究近年来的相关论文,归纳上述问题的有效解决办法,并且对同一类问题的解决办法进行对比分析,梳理出不同场景下的有效处理策略,同时为优化模型提供一定的解决思路。
In the field of target detection based on deep learning,with the progress of image feature extraction technology,anchor frame can be used to generate frames in different positions of the image,and the frame position can be regressed by extracting the border region features.In the actual training process,it is found that the single-stage detection method based on anchor has insufficient accuracy.After analysis and summary,the problems are divided into three categories:unbalanced sample classification;multi-scale target detection;inconsistent optimization target and inference.After studying the related papers in recent years,this paper summarizes the effective solutions to the above problems,compares and analyzes the solutions to the same kind of problems,sorts out the effective processing strategies in different scenarios,and provides certain solutions for the optimization model.
作者
单莉
梁煜博
SHAN Li;LIANG Yubo(Northern Theater Naval Staff,Qingdao Shandong 266000,China;Beijing Union University,Beijing 100101,China)
出处
《北京工业职业技术学院学报》
2020年第4期10-15,共6页
Journal of Beijing Polytechnic College
关键词
深度学习
目标检测
单阶段检测算法
基于锚框算法
精度提升
deep learning
target detection
single-stage detection
anchor-based algorithm
accuracy upgrade
作者简介
单莉(1981—),女,山东青岛人,工程师,工学硕士,研究方向为计算机技术与网络信息安全。
