基于多尺度重叠滑动池化的SSD果冻杂质检测方法  

Jelly Impurity Detection Based on Scalable Overlapping Slide Pooling SSD Network

作　　者：张芳慧 章春娥[1,2] 张琳娜 岑翼刚[1,2] 阚世超[1,2] Zhang Fanghui;Zhang Chune;Zhang Linna;Cen Yigang;Kan Shichao(Institute of Information Science,Beijing Jiaotong University,Beijing 100044,China;Beijing Key Laboratory of Advanced Information Science and Network Technology,Beijing 100044,China;School of Mechanical Engineering,Guizhou University,Guiyang,Guizhou 550025,China)

机构地区：[1]北京交通大学信息科学研究所,北京100044 [2]现代信息科学与网络技术北京市重点实验室,北京100044 [3]贵州大学机械工程学院,贵州贵阳550025

出　　处：《信号处理》2020年第11期1811-1818,共8页Journal of Signal Processing

基　　金：中央高校基本科研业务费专项资金资助(2019YJS046);贵州省自然科学基金(黔科合基础[2019]1064);国家自然科学基金(62062021,61872034);北京市自然科学基金(4202055)资助。

摘　　要：提出一种基于SSD的杂质检测方法,用于检测生产线中果冻内部的杂质,并标注出杂质的类型和位置。在预处理阶段,提出滑动图像块分割方法,将整张果冻图像分割成若干图像块,避免杂质占比过小,造成准确率低的现象。使用迁移学习的方法,将神经网络在ImageNet数据库上学习到的特征迁移到果冻数据库中,加快网络收敛速度,同时,在一定程度上避免了过拟合现象。提出多尺度重叠滑动池化(SOSP)方法,取代第五层池化以取得更加鲁棒的特征池化。最后,将一幅图下的所有分割块上的检测结果进行整合,得到整张图像的检测结果。实验结果表明,本文提出的方法有效可行,对多种缺陷平均准确率达到0.7271。相比其他方法,本文的算法更具鲁棒性,可应用到果冻生产线中。An impurity detection method based on SSD is proposed to detect the internal impurity of jelly in the production line.In the pre-processing stage,a sliding image block segmentation method is proposed to segment the whole jelly image into several segmentation block images,so as to avoid the low accuracy because the image is too large and the target is too small.Using the method of transfer learning,the features learned from the neural network on ImageNet are transferred to the jelly database to accelerate the convergence speed of the network and avoid over-fitting to a certain extent.A scalable overlapping slide pooling(SOSP)method is proposed to replace the fifth pooling to achieve more robust feature pooling.Finally,the detection results on all segmentation block images under one image are integrated to obtain the detection results of the whole image.Experimental results show that the proposed method is effective and feasible.The average accuracy for the multiple defects reaches 0.7271.Compared with other methods,our algorithm is more robust and can be applied to the jelly impurity detection production line.

关 键 词：果冻杂质检测 深度学习 多尺度重叠滑动池化 迁移学习 目标检测算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]