面向单一背景的改进RetinaNet目标检测方法研究  

Study on Single Background Object Detection Oriented Improved-RetinaNet Model and Its Application

作　　者：周波 江佩峰[1,2] 段昶 罗月童 ZHOU Bo;JIANG Peifeng;DUAN Chang;LUO Yuetong(School of Computer Science and Information Engineering,Hefei University of Technology,Hefei 230601,China;Engineering Research Center of Safety Critical Industrial Measurement and Control Technology,Ministry of Education,Hefei 230009,China)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601 [2]安全关键工业测控技术教育部工程研究中心,合肥230009

出　　处：《计算机科学》2023年第7期137-142,共6页Computer Science

基　　金：国家自然科学基金(61602146);国家重点基础研究发展计划(2017YFB1402200);安徽省科技攻关计划(1604d0802009)。

摘　　要：基于深度学习的目标检测算法在工业缺陷检测领域得到了充分推广与应用,但少有适用于工业检测场景中单一背景的算法。文中以工业检测场景中具有大量简单重复背景为出发点,对RetinaNet算法进行了如下改进:1)引入难负样本挖掘策略,减小了大量简单重复负样本对对模型拟合正样本的影响;2)设计了自适应忽略样本选择策略,避免与正样本交并比高的样本混入负样本而混淆模型训练;3)简化了RetinaNet的分类子网络,降低了模型改进后的过拟合风险。在公开的PCB缺失孔数据集及自建的LED气泡数据集上,相比RetinaNet算法,改进后的方法在召回率上分别提升了14.1%和1.8%,在精确率上分别提升了3.6%和0.4%,表明改进后的方法能显著提升单一背景下的目标检测水平。Object detection algorithms based on deep learning have been fully promoted and applied in the field of industrial defect detection,but few algorithms are suitable for single background in industrial detection scenarios.This paper takes the industrial detection scene with a large number of simple repeated backgrounds as the starting point,and makes the following improvements to the RetinaNet algorithm:1)introduce the difficult negative sample mining strategy to reduce the impact of a large number of simple repeated negative samples on the model fitting positive samples;2)an adaptive ignoring sample selection strategy is designed to avoid mixing samples with high intersection ratios of positive samples into negative samples to confuse model training;3)the classification sub-network of RetinaNet is simplified,and the risk of overfitting after model improvement is reduced.Compared with the RetinaNet algorithm,the improved method improves the recall rate by 14.1%and 1.8%,and the precision rate by 3.6%and 0.4%respectively on the public PCB missing hole dataset and the self-built LED bubble dataset,indicating that the improved method can significantly improve the level of object detection in a single background.

关 键 词：深度学习 RetinaNet 自适应采样 单一背景 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]