CASIA OpenIR  > 毕业生  > 博士学位论文
复杂纹理玻璃容器内液体杂质视觉检测方法研究
郭跃1,2
学位类型工学博士
导师原魁
2018-05-27
学位授予单位中国科学院研究生院
学位授予地点北京
关键词杂质检测 复杂纹理玻璃容器 区域生成 端到端 关联样本 序列区域
其他摘要
    液体商品广泛出现在人们的日常生活中,从医用的药剂到饮料等,它们的质量直接关系到饮用者的生理健康。随着社会经济的发展与人民生活质量的改善,人们在面对饮料或酒类等液体商品时选择更多。通常在消费者眼里,精致的包装意味着该商品的档次很高。因此,以不透明容器为包装的液体商品在市场上占有很大的比重。为了确保不透明容器内液体的质量,除了从源头避免杂质进入液体商品以外,对不透明容器内的杂质进行检测不可或缺,并需要耗费大量的人力,检查纹理更加复杂的容器也会增加人工灯检的负担。
    本文根据相关企业的需求,以不透明玻璃瓶体为主的复杂纹理玻璃容器为对象,探讨了不透明玻璃容器内液体杂质检测方法:借鉴了透明玻璃容器内液体杂质检测的思路,提出并分析了复杂纹理玻璃容器内液体杂质检测问题,通过不同角度的实验验证了复杂纹理玻璃容器内液体杂质检测方法的可行性和有效性。本论文的主要工作与创新点包括以下几个方面:
    1、介绍了目标检测的研究现状,分别给出了透明玻璃容器内液体杂质检测和复杂纹理玻璃容器内液体杂质检测的研究状况,介绍了本文的前期工作基础以及本文拟解决的主要问题,并给出了本文的研究内容和基本结构。
    2、给出了基于区域的杂质检测。根据复杂纹理玻璃容器的特点,提出了一种基于纹理先验特征的区域生成方法。为了合并聚集在图像局部空间中的小区域,结合了选择性搜索或基于超像素的区域生成方法。在不排除过多包含杂质的生成区域的前提下,使用基于纹理先验特征和超像素的区域生成方法能够输出很少的生成区域,从而极大地减少了区域分类模型的计算量。
    3、设计了端到端的杂质检测。分离了灰度图像和差分图像的特征提取路径,分别将常用于语义分割任务的全卷积网络和U型网络扩展成孪生全卷积网络和孪生U型网络。相比全卷积网络和U型网络,孪生全卷积网络和孪生U型网络在杂质检测任务上的表现更好。与使用大规模数据训练并通过小数据微调的Faster R-CNN相比,使用小规模数据重新训练得到的孪生全卷积网络的检测结果接近前者的结果。
    4、提出了基于关联区域的杂质检测。提出了一种基于关联样本的数据增广方法,训练了关联区域分类模型,并在此基础上提出了基于分类-关联的集成杂质检测方法。为了验证基于关联样本的数据增广方法的有效性,分别将原有样本和用生成对抗网络(WGAN-GP)合成的新样本、对抗样本和关联样本结合来训练区域分类模型。实验结果表明,使用基于关联样本的数据增广方法能够进一步改善区域分类的表现,而使用基于分类-关联的集成杂质检测方法则能够在弱曝光情况下检测人眼也难以分辨的难例。
    5、提出了基于序列区域的杂质检测。针对运动杂质的特点,提出了一种基于图像特征和运动特征的空间注意力模型,构建了基于长时循环卷积网络的序列区域分类模型,并在此基础上提出了一种基于序列区域的集成杂质检测方法。实验结果表明,使用本章提出的检测方法能够得到比较理想的杂质检测效果,为复杂纹理玻璃容器内杂质检测的实际应用提供了可能性。
;
    Liquid products exist everywhere in people's daily lives, from the medical drugs to some other liquid necessities, and their qualities are directly related to drinkers' physical health. The development of the social economy and the improvement of life quality leave people overwhelming choices when facing liquid products such as drinks or alcohols, and consumers usually think that liquid in an exquisite bottle tends to be fine. To ensure the qualities of opaque liquid products, producers not only have to avoid impurities flowing in the liquid from the source but also have to detect potential impurities inside the products. A large number of human checkers are needed for manual impurity detection tasks, what's worse, more complicated bottles aggravated the difficulties of artificial light inspections. 
    This dissertation investigates the techniques of impurity detection in containers with complex textures such as opaque glass bottles: inspired by the relevant techniques of impurity detection in transparent bottles, the impurity detection problem in opaque glass bottles is proposed and analysed, experimental results with different impurity detection architectures show the feasibility and the effectiveness of impurity detection in opaque glass bottles. Major works and innovations of this dissertation are as follows:
    1. The research results of object detection are briefly introduced, and current developments about impurity detection in transparent containers and that in opaque glass containers with complex textures are then provided respectively, then previous works of our team and major problems we aim to solve are introduced, finally, the research contents and basic structures of this dissertation is introduced.
    2. Region-based impurity detection architectures are built. A region proposal generation method based on texture priors is proposed according to the characteristics of glass containers with complex textures. To further combine the clusters of tiny region proposals, selective search or region proposal methods based on superpixels are integrated into our method. Only a few proposals are provided using region proposal methods based on texture priors and superpixels without losing many outputs with true impurities, and it greatly reduces the computational costs of region classification models. 
    3. End-to-end impurity detection models are built. Feature extractions from gray images and differential ones are separated, and fully convolutional neural networks and u-nets are extended into siamese fully convolutional neural networks and siamese u-nets, and they perform better than traditional fully convolutional neural networks and u-nets. Finally, a fully convolutional neural network retrained with small datasets performs similarly with Faster R-CNN finetuned with small datasets.
    4. Correlational region based impurity detection architectures are proposed. Specifically, a data augmentation method based on classification and correlation is proposed. To evaluate the effectiveness of data augmentation methods for correlational examples, new datasets combined with synthetic patches using generative adversarial networks (WGAN-GP), those with adversarial examples, and those with correlational examples are used to train region classification models. Extensive experiments demonstrate that such data augmentation method can be used to further improve the region classification performances, and with correlational examples, an ensemble impurity detection method can detect hard examples under inadequate exposures. 
    5. Sequential region based impurity detection architectures are proposed. According to the characteristics of moving impurities, a spatial attention model based on features of vision and motion is proposed, then a sequential region classification model using a long-term recurrent convolutional network is proposed. Experimental results show that detection performances are the state-of-the-art using architectures in this chapter, which provides the possibilities of applying automatic impurity detection in glass containers with complex textures.
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20957
专题毕业生_博士学位论文
作者单位1.中国科学院自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
郭跃. 复杂纹理玻璃容器内液体杂质视觉检测方法研究[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
郭跃 - 论文 - 签字版.pdf(8075KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[郭跃]的文章
百度学术
百度学术中相似的文章
[郭跃]的文章
必应学术
必应学术中相似的文章
[郭跃]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。