基于RetinaNet的密集型钢筋计数改进算法  

Improved counting algorithm for dense rebars based on RetinaNet

作　　者：明洪宇 陈春梅 刘桂华 邓豪[1,2] MING Hongyu;CHEN Chunmei;LIU Guihua;DENG Hao(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China;Robot Technology Used for Special Environment Key Laboratory of Sichuan Province,Mianyang 621010,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]特殊环境机器人技术四川省重点实验室,四川绵阳621010

出　　处：《传感器与微系统》2020年第12期115-118,共4页Transducer and Microsystem Technologies

基　　金：国防科工局核能开发科研项目([2016]1295);四川省科技厅重点研发资助项目(19ZS2117)。

摘　　要：提出了一种基于RetinaNet目标检测框架,结合高斯混合模型(GMM)和期望最大化(EM)算法的钢筋计数方法。通过在RetinaNet特征提取后端增加Soft-IOU层以对预测框与真实框的交并比进行评估。借助Soft-IOU评估到的质量分数,生成钢筋目标检测的高斯混合模型。针对RetinaNet原始框架对密集目标检测效果欠理想的问题,采用了基于EM算法的高斯混合聚类方法解决歧义检测以提高计数精度。实验结果表明:改进后的方法较RetinaNet算法平均精度提高了3.3%,计数均方根误差提升了64.2,具有很强的适应性。An object detection framework based on RetinaNet is proposed by using Gaussian mixture model(GMM)and expectation maximization(EM).Firstly,a Soft-IOU layer is added to the back end of RetinaNet’s feature extraction to evaluate the intersection over union(IOU)of the predicted bounding boxes and the ground truth.Then,with the help of the quality score evaluated by Soft-IOU,the GMM for rebars detection is generated.Finally,in order to improve the counting precision,Gaussian mixture clustering method based on EM algorithm is used to solve the problem that the original RetinaNet framework is not ideal for dense object detection.The experimental results show that the average precision of the improved method is 3.3%higher than that of the RetinaNet,and the counting root mean square error is improved by 64.2.Therefore,the method shows strong adaptability.

关 键 词：RetinaNet网络 期望最大化(EM)算法 钢筋计数 高斯混合模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]