【作者】 陈震；

【导师】 王鸿鹏；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 缺陷检测是工业生产领域的一个关键环节,旨在发现各种工业制成品的外部及内部瑕疵,是使产品质量满足客户要求的重要技术之一。现实工业生产中每个产品都有各自不同的大小型号,同时由于缺陷检测要求产品需要具备不同角度的拍摄照片以方便检测。为了训练出符合生产要求的缺陷检测模型,通常需要大规模且高质量的训练数据集。一些研究者发现,在监督学习问题中,存在标记成本较为昂贵且标记难以大量获取的问题。针对一些特定任务,只有行业专家才能为样本做上准确标记,在此背景下,训练出优秀的缺陷检测模型需要我们标记更多的数据。与此同时,另一部分研究者发现,模型在学习学习新的任务时会出现灾难性遗忘遗忘问题,从而遗忘旧类别的知识。联系到缺陷检测中,由于下游检测产线通常缺少模型训练原有的旧类别样本数据集,没有条件对模型做全量训练,这对模型性能提升有很大的限制。为了减轻产线工人的标记任务工作量,本文提出了结合主动学习的电容缺陷样本数据集标注策略,旨在通过筛选出更具备代表性的样本,模型通过在被筛选出的样本组成的训练集上训练,进而使得模型的性能收益最大化。为了实现此目标,本文提出了一种:在预处理阶段,候选样本通过数据增强生成了多个patch;本文通过利用各个patch在模型预测上的一致性对候选样本进行打分,通过对候选样本集的样本进行分数排序选择分数最高的样本,即最具代表性的样本。为了验证所提出的策略的性能,本文基于真实的薄膜电容数据集进行了实验研究,实验结果证明,本文的策略可以自动筛选出更有价值的数据,且在筛选出的数据组成的数据集上可以保持样本间的类别平衡。为了解决灾难性遗忘对检测模型的影响,本文提出了一种将开集识别与增强判别模块结合的持续学习方法,在缺少旧类别样本的情况下,使得模型保留旧类别样本知识的同时尽可能得将新旧类别样本区分开。为了帮助模型在持续学习过程中更多地识别出旧类别样本,本文设计了开集识别模块。为了有效地克服灾难性遗忘,本文通过构建增强判别模块,充分利用了新旧类别样本来帮助模型克服灾难性遗忘。为了验证所提出持续学习方法的性能,本文在CIFAR-100,Image Net-100以及基于真实薄膜电容数据集进行了对比实验,实验结果表明,本文的方法可以在缺少旧类别样本的情况下有效克服灾难性遗忘。更多还原

【Abstract】 Defect detection is a key aspect of industrial production,aiming to find external and internal defects in various industrial manufactured products,and is one of the most important technologies to make product quality meet customer requirements.The reality of industrial production is that each product has its own size and type,and because defect detection requires that the product be photographed from different angles to facilitate detection.In order to train a defect detection model that meets the production requirements,a large and high quality training data set is usually required.Some researchers have found that in supervised learning problems,markers are expensive and difficult to obtain in large quantities.For some specific tasks,only industry experts can accurately label the samples,and in this context,training good defect detection models requires us to label more data.At the same time,another part of the researchers found that the models suffer from a catastrophic forgetting problem when learning to learn new tasks,thus forgetting the knowledge of old categories.In connection to defect detection,there is no condition to do full training on the model because the downstream inspection line usually lacks the original old category sample dataset for model training,which has a great limitation on the model performance improvement.In order to reduce the marking task workload of production line workers,this paper proposes a capacitive defect sample dataset labeling strategy combined with active learning,aiming to maximize the performance gain of the model by screening out more representative samples and training the model on the training set composed of the screened samples.To achieve this goal,this paper proposes a strategy in which: in the preprocessing stage,candidate samples are generated by data augmentation with multiple patches;this paper scores the candidate samples by using the consistency of each patch in model prediction,and selects the sample with the highest score,i.e.,the most representative sample,by ranking the samples in the candidate sample set by their scores.In order to verify the performance of the proposed strategy,this paper conducts an experimental study based on a real thin film capacitance dataset,and the experimental results prove that the strategy in this paper can automatically filter out more valuable data,and the category balance among samples can be maintained on the dataset composed of the filtered data.In order to address the impact of catastrophic forgetting on the detection model,this paper proposes a continual learning method that combines open-set recognition with an enhanced discriminant module to make the model retain the knowledge of the old category samples while distinguishing the old and new category samples as much as possible in the absence of the old category samples.In order to help the model identify more old category samples in the continual learning process,an open-set recognition module is designed in this paper.In order to effectively overcome catastrophic forgetting,this paper builds an enhanced discrimination module to make full use of the old and new category samples to help the model overcome catastrophic forgetting.In order to verify the performance of the proposed continual learning method,comparative experiments are conducted in CIFAR-100,Image Net-100 and real film capacitance-based datasets,and the experimental results show that the method in this paper can effectively overcome catastrophic forgetting in the absence of old category samples.更多还原