适于少样本缺陷检测的两阶段缺陷增强网络

缺陷检测模型一般需要大量样本来学习缺陷的特征,但实际场景中一些重要缺陷的样本难以收集,如何用少量样本来学习罕见缺陷的特征成为一个具有挑战性的问题。为了促进少样本缺陷检测的研究,构建了一个新的工业表面缺陷数据集,包括缺陷样本和无缺陷样本。同时提出了一个两阶段缺陷增强网络以提升少样本场景下的缺陷检测性能,它利用了无缺陷样本,并将整个训练过程分为两个阶段。第一阶段的训练需要大量缺陷样本,而第二阶段的训练只需要少量缺陷样本和无缺陷样本。此外,还提出了一个缺陷突显模块,可以更好地利用无缺陷样本来增强缺陷区域的特征。在新数据集上的实验表明,该缺陷检测模型的性能优于其他的少样本目标检测模型,在工业表面缺陷检测中具有更好的应用前景。     

A fast regularity measure for surface defect detection

In this paper, we propose a fast regularity measure for defect detection in non-textured and homogeneously textured surfaces, with specific emphasis on ill-defined subtle defects. A small neighborhood window of proper size is first chosen and they slide over the entire inspection image in a pixel-by-pixel basis. The regularity measure for each image patch enclosed in the window is then derived from the eigenvalues of the covariance matrix formed by the variance–covariance of the x- and y-coordinates with the pixel gray levels as the weights for all pixel points in the window. The two eigenvalues of the weighted covariance matrix will be approximately the same when the image patch contains only a homogeneous region, whereas the two eigenvalues will be relatively different if the image patch in the window contains a defect. The smaller eigenvalue of the covariance matrix is then used as the regularity measure. The integral image technique is introduced to the computation of the regularity measure so that it is invariant to the neighborhood window size. The proposed method uses only one single discrimination feature for defect detection. It avoids the use of complicated classifiers in a high-dimensional feature space, and requires no learning process from a set of defective and defect-free training samples. Experimental results on a variety of material surfaces found in industry, including textured images of plastic surfaces and leather and non-textured images of backside solar wafers and LCD backlight panels, have shown the effectiveness of the proposed regularity measure for surface defect detection. It is computationally very fast, and takes only 0.032 s for a 400 × 400 image on a Pentium 3.00?GHz personal computer. In a test set of 73 backside solar wafer images involving 53 defect-free and 20 defective samples, the proposed regularity measure can correctly identify all the test images.     

Defect detection in multi-crystal solar cells using clustering with uniformity measures

Solar cells that convert sunlight into electrical energy are the main component of a solar power system. Quality inspection of solar cells ensures high energy conversion efficiency of the product. The surface of a multi-crystal solar wafer shows multiple crystal grains of random shapes and sizes. It creates an inhomogeneous texture in the surface, and makes the defect inspection task extremely difficult. This paper proposes an automatic defect detection scheme based on Haar-like feature extraction and a new clustering technique. Only defect-free images are used as training samples. In the training process, a binary-tree clustering method is proposed to partition defect-free samples that involve tens of groups. A uniformity measure based on principal component analysis is evaluated for each cluster. In each partition level, the current cluster with the worst uniformity of inter-sample distances is separated into two new clusters using the Fuzzy C-means. In the inspection process, the distance from a test data point to each individual cluster centroid is computed to measure the evidence of a defect. Experimental results have shown that the proposed method is effective and efficient to detect various defects in solar cells. It has shown a very good detection rate, and the computation time is only 0.1 s for a 550 × 550 image.     

基于轻量化重构网络的表面缺陷视觉检测

基于深度学习的方法在某些工业产品的表面缺陷识别和分类方面表现出优异的性能,然而大多数工业产品缺陷样本稀缺,而且特征差异大,导致这类需要大量缺陷样本训练的检测方法难以适用.提出一种基于重构网络的无监督缺陷检测算法,仅使用容易大量获得的无缺陷样本数据实现对异常缺陷的检测.提出的算法包括两个阶段:图像重构网络训练阶段和表面缺陷区域检测阶段.训练阶段通过一种轻量化结构的全卷积自编码器设计重构网络,仅使用少量正常样本进行训练,使得重构网络能够生成无缺陷重构图像,进一步提出一种结合结构性损失和L1损失的函数作为重构网络的损失函数,解决自编码器检测算法对不规则纹理表面缺陷检测效果较差的问题;缺陷检测阶段以重构图像与待测图像的残差作为缺陷的可能区域,通过常规图像操作即可实现缺陷的定位.对所提出的重构网络的无监督缺陷检测算法的网络结构、训练像素块大小、损失函数系数等影响因素进行了详细的实验分析,并在多个缺陷图像样本集上与其他同类算法做了对比,结果表明重构网络的无监督缺陷检测算法有较强的鲁棒性和准确性.由于重构网络的无监督缺陷检测算法的轻量化结构,检测1 024×1 024像素图像仅仅耗时2.82 ms,...     

软件缺陷预测中的数据预处理方法

软件缺陷预测是软件质量保障领域的热点研究课题,缺陷预测模型的质量与训练数据有密切关系。用于缺陷预测的数据集主要存在数据特征的选择和数据类不平衡问题。针对数据特征选择问题,采用软件开发常用的过程特征和新提出的扩展过程特征,然后采用基于聚类分析的特征选择算法进行特征选择;针对数据类不平衡问题,提出改进的Borderline-SMOTE过采样方法,使得训练数据集的正负样本数量相对平衡且合成样本的特征更符合实际样本特征。采用bugzilla、jUnit等项目的开源数据集进行实验,结果表明：所采用的特征选择算法在保证模型F-measure值的同时,可以降低57.94%的模型训练时间;使用改进的Borderline-SMOTE方法处理样本得到的缺陷预测模型在Precision、Recall、F-measure、AUC指标上比原始方法得到的模型平均分别提高了2.36个百分点、1.8个百分点、2.13个百分点、2.36个百分点;引入了扩展过程特征得到的缺陷预测模型比未引入扩展过程特征得到的模型在F-measure值上平均提高了3.79%;与文献中的方法得到的模型相比,所提方法得到的模型在F-measure值上平均提高了15.79%。实验结果证明所提方法能有效提升缺陷预测模型的质量。     

软件缺陷预测中的数据预处理方法

软件缺陷预测是软件质量保障领域的热点研究课题，缺陷预测模型的质量与训练数据有密切关系。用于缺陷预测的数据集主要存在数据特征的选择和数据类不平衡问题。针对数据特征选择问题，采用软件开发常用的过程特征和新提出的扩展过程特征，然后采用基于聚类分析的特征选择算法进行特征选择；针对数据类不平衡问题，提出改进的Borderline-SMOTE过采样方法，使得训练数据集的正负样本数量相对平衡且合成样本的特征更符合实际样本特征。采用bugzilla、jUnit等项目的开源数据集进行实验，结果表明：所采用的特征选择算法在保证模型F-measure值的同时，可以降低57.94%的模型训练时间；使用改进的Borderline-SMOTE方法处理样本得到的缺陷预测模型在Precision、Recall、F-measure、AUC指标上比原始方法得到的模型平均分别提高了2.36个百分点、1.8个百分点、2.13个百分点、2.36个百分点；引入了扩展过程特征得到的缺陷预测模型比未引入扩展过程特征得到的模型在F-measure值上平均提高了3.79%；与文献中的方法得到的模型相比，所提方法得到的模型在F-measure值上平均提高了15.79%。实验结果证明所提方法能有效提升缺陷预测模型的质量。     

Defect detection of uneven brightness in low-contrast images using basis image representation

In this paper, we propose a machine vision approach for detecting local irregular brightness in low-contrast surface images and, especially, focus on mura (brightness non-uniformity) defects in liquid crystal display (LCD) panels. A mura defect embedded in a low-contrast surface image shows no distinct intensity from its surrounding region, and even worse, the sensed image may also present uneven illumination on the surface. All these make the mura defect detection in low-contrast surface images extremely difficult.A set of basis images derived from defect-free surface images are used to represent the general appearance of a clear surface. An image to be inspected is then constructed as a linear combination of the basis images, and the coefficients of the combination form the feature vector for discriminating mura defects from clear surfaces. In order to find minimum number of basis images for efficient and effective representation, the basis images are designed such that they are both statistically independent and spatially exclusive. An independent component analysis-based model that finds both the maximum negentropy for statistical independency and minimum spatial correlation for spatial redundancy is proposed to extract the representative basis images. Experimental results have shown that the proposed method can effectively detect various mura defects in low-contrast LCD panel images.     

基于卷积自编码器的医用玻璃瓶口缺陷检测方法

为解决缺陷检测中缺陷样本数量少、种类多、难以提供足够的数据来进行有监督深度学习模型训练的问题,本文利用工业生产中大量易获取没有缺陷的正样本数据,建立Encoder-Decoder结构的卷积自编码网络缺陷检测模型,将空间和通道注意力的卷积注意力模块嵌入编码器中增强网络特征提取能力。在编码阶段加入上下文信息模块,获得更大的感受野,减小计算量。同时,结合多尺度结构相似性MS-SSIM和L1损失来改善图像重构效果,使用峰值信噪比PSNR衡量重构误差并判别异常。实验结果表明,提出的医用玻璃瓶口缺陷检测方法能够准确检出缺陷数据和分割缺陷区域,精确度为99.45%、召回率为97.63%、漏检率为0.55%、误检率为2.93%。该方法能够准确检出玻璃瓶口缺陷,定位缺陷区域,同时图像重构耗时短,仅需10.37 ms左右,能够实现准确、高效的自动化产品质量检测。     

A new Feature-Fusion method based on training dataset prototype for surface defect recognition

Surface defect recognition is important to improve the surface quality of end products. In this area, there were many convolutional neural network (CNN)-based methods because CNN can extract features automatically. The extracted features determine the performance of recognition, so it is important for CNN-based methods to extract effective and sufficient features. However, feature extraction needs a large-scale dataset, which is hard to obtain. To save the cost of collecting samples and extract effective features, ensemble methods were proposed to make full use of the features extracted by CNN in order to guarantee good performance with limited samples. However, the methods are confined to utilize one sample – they extracted multi-level features from one individual sample – but ignore the vast information in a dataset. Due to the limit information in one sample, this paper turns the attention to the training dataset and attempts to mine the multi-level information in the dataset for predicting. The proposed method is named as Prototype vectors fusion-based CNN (ProtoCNN), which utilizes the prototype information in the training dataset. In training process, it trains a VGG11 as the base model, and meanwhile prototype vectors corresponding to each defect class are generated in multiple feature layers of VGG11. Then, in predicting process, the prototype vectors are fused to predict unknown samples. The experiments on three famous datasets, including NEU-CLS, wood dataset, and textile dataset indicate that the proposed ProtoCNN outperforms conventional ensemble models and other models for surface defect recognition. In these datasets, ProtoCNN has achieved the accuracy of 99.86%, 90.01%, and 81.28% respectively, which increase 1.05%, 4.07%, 19.53% compared to its base model respectively. Finally, this paper analyzes the effectiveness and practicality of prototype vectors, showing that the proposed ProtoCNN is practical for real world application.     

Improvement of the kernel minimum squared error model for fast feature extraction

The kernel minimum squared error (KMSE) expresses the feature extractor as a linear combination of all the training samples in the high-dimensional kernel space. To extract a feature from a sample, KMSE should calculate as many kernel functions as the training samples. Thus, the computational efficiency of the KMSE-based feature extraction procedure is inversely proportional to the size of the training sample set. In this paper, we propose an efficient kernel minimum squared error (EKMSE) model for two-class classification. The proposed EKMSE expresses each feature extractor as a linear combination of nodes, which are a small portion of the training samples. To extract a feature from a sample, EKMSE only needs to calculate as many kernel functions as the nodes. As the nodes are commonly much fewer than the training samples, EKMSE is much faster than KMSE in feature extraction. The EKMSE can achieve the same training accuracy as the standard KMSE. Also, EKMSE avoids the overfitting problem. We implement the EKMSE model using two algorithms. Experimental results show the feasibility of the EKMSE model.     

A CAD/CAE incorporate software framework using a unified representation architecture

Currently, there are still some big gaps between the CAD system and CAE system, e.g. the different data structure for model representation, which costs lots of time and effort of engineers in the interaction between these two kinds of systems. In order to bridge these gaps, an incorporate software framework is proposed in this paper. In this framework, the unified representation architecture (URA) is presented that makes CAD and CAE to be an organic entity. The URA contains three components: (1) unified data model (UDD) including unified B-rep, unified feature and unified mesh; (2) unified data management (UDM) consisting of unified interaction, unified data structure, unified Constructive Solid Geometry (CSG) history and unified interface; (3) unified display and post-processor (UDP) for both design and performance analysis. The URA facilitates the incorporation by explicitly representing design and analysis information as design features, which maintains their associations through the history chain. Besides the URA, a unified mesh data (UMD) is proposed to unify the mesh of CAD model display and CAE analysis with the purpose of reducing the redundancy of mesh data. The unified mesh data (UMD) is proposed to unify the mesh of CAD model display and CAE analysis, which greatly reduces the redundancy of mesh generation data. Finally, the high efficiency of the proposed framework is demonstrated by engineering examples.     

An intelligent fault diagnosis method for rotor-bearing system using small labeled infrared thermal images and enhanced CNN transferred from CAE

Despite deep learning models can largely release the pressure of manual feature engineering in intelligent fault diagnosis of rotor-bearing systems, their performance mostly depends on enough labeled samples constructed from the vibration signals. Acquiring lots of labeled samples is often laborious, and the vibration sensors tightly fixed on the equipment may influence their structures after long time running. To address these two problems, a new framework based on small labeled infrared thermal images and enhanced convolutional neural network (ECNN) transferred from convolutional auto-encoder (CAE) is proposed. First, infrared thermal images are measured to characterize various health states of rotor-bearing system. Second, exponential linear unit (ELU) and stochastic pooling (SP) are used to construct ECNN. Then, the model parameters of a CAE pre-trained with unlabeled thermal images are transferred to initialize the ECNN. Finally, small labeled thermal images are used for training ECNN to further adjust model parameters. The collected thermal images are used to test the diagnosis performance of the proposed method. The analysis and comparison results show that the proposed method outperforms the current mainstream methods.     

基于堆叠降噪自动编码器的胶囊缺陷检测方法

目前医用胶囊生产过程中的缺陷检测主要由人工完成,费时费力,容易受主观因素的影响。提出一种基于堆叠降噪自动编码器的胶囊表面缺陷检测方法,该方法首先建立深度自动编码器网络,并根据缺陷样本进行降噪训练,获取网络的初始权值;然后通过BP算法进行微调,得到训练样本到无缺陷模板之间的映射关系;最后利用重构图像与缺陷图像之间的对比关系,实现测试样本的缺陷检测。实验表明,堆叠降噪自动编码器较好地建立了上述映射关系,能快速、准确地进行缺陷检测,对噪声具有很强的鲁棒性和稳定性。     

Multi-Resolution and Noise-Resistant Surface Defect Detection Approach Using New Version of Local Binary Patterns

Visual quality inspection systems play an important role in many industrial applications. In this respect, surface defect detection is one of the problems that have received much attention by image processing scientists. Until now, different methods have been proposed based on texture analysis. An operation that provides discriminate features for texture analysis is local binary patterns (LBP). LBP was first introduced for gray-level images that makes it useless for colorful samples. Sensitivity to noise is another limitation of LBP. In this article, a new noise-resistant and multi-resolution version of LBP is used that extracts color and texture features jointly. Then, a robust algorithm is proposed for detecting abnormalities in surfaces. It includes two steps. First, new version of LBP is applied on full defect-less surface images, and the basic feature vector is calculated. Then, by image windowing and computing the non-similarity amount between windows and basic vector, a threshold is computed. In test phase, defect parts are detected on test samples using the tuned threshold. High detection rate, low computational complexity, low noise sensitivity, and rotation invariant are some advantages of our proposed approach.     

Design of multi-scale receptive field convolutional neural network for surface inspection of hot rolled steels

Due to the large intra-class variations and unbalanced training samples, the accuracy of existing algorithms used in defect classification of hot rolled steels is unsatisfactory. In this paper, a new hierarchical learning framework is proposed based on convolutional neural networks to classify hot rolled defects. Multi-scale receptive field is introduced in the new framework to extract multi-scale features, which can better represent defects than the feature maps produced by a single convolutional layer. A group of AutoEncoders are trained to reduce the dimension of the extracted multi-scale features which improve the generalization ability under insufficient training samples. Besides, to mitigate the deviation caused by fine-tuning the pre-trained model with images of different context, we add a penalty term in the loss function, which is to reconstruct the input image from the feature maps produced by the pre-trained model, to help network encode more effective and structured information. The experiments with samples captured from two hot rolled production lines showed that the proposed framework achieved a classification rate of 97.2% and 97% respectively, which are much higher than the conventional methods.     

A semi-supervised convolutional neural network-based method for steel surface defect recognition

Automatic defect recognition is one of the research hotspots in steel production, but most of the current methods focus on supervised learning, which relies on large-scale labeled samples. In some real-world cases, it is difficult to collect and label enough samples for model training, and this might impede the application of most current works. The semi-supervised learning, using both labeled and unlabeled samples for model training, can overcome this problem well. In this paper, a semi-supervised learning method using the convolutional neural network (CNN) is proposed for steel surface defect recognition. The proposed method requires fewer labeled samples, and the unlabeled data can be used to help training. And, the CNN is improved by Pseudo-Label. The experimental results on a benchmark dataset of steel surface defect recognition indicate that the proposed method can achieve good performances with limited labeled data, which achieves an accuracy of 90.7% with 17.53% improvement. Furthermore, the proposed method has been applied to a real-world case from a Chinese steel company, and obtains an accuracy of 86.72% which significantly better than the original method in this workshop.     

A CAD-CAE Integrated Injection Molding Design System

. In the injection molding design process, interaction between design and analysis is very intensive. This is to ensure that the plastic part being designed is manufacturable by the injection molding process. However, such interaction is not supported by current computer-aided systems (CAD and CAE), because design and analysis are realized as isolated modules. Although most of CAE systems provide built-in modeling tools, these are only meant for developing an analysis model with very limited CAD functionality. On the other hand, some CAD systems have allowed certain CAE systems to run under their environments, but inherently they use different data models, thus communication between them is poor. This paper presents an innovative, CAD-CAE integrated, injection molding design system. This system uses an integrated data model for both design and analysis. The system is built on top of existing CAD and CAE systems, which not only greatly saves development effort, but also makes full use of the strong functionality of commercial computer aided systems. The system architecture consists of four layers: a CAD and CAE platform layer; a CAD-CAE feature layer; a model layer; and a GUI layer. Two design cases were studied to illustrate the iterative design-analysis process and use of the developed system.     

基于深度自编码网络的软件缺陷预测方法

软件缺陷预测是提升软件质量的有效方法,而软件缺陷预测方法的预测效果与数据集自身的特点有着密切的相关性。针对软件缺陷预测中数据集特征信息冗余、维度过大的问题,结合深度学习对数据特征强大的学习能力,提出了一种基于深度自编码网络的软件缺陷预测方法。该方法首先使用一种基于无监督学习的采样方法对6个开源项目数据集进行采样,解决了数据集中类不平衡问题;然后训练出一个深度自编码网络模型。该模型能对数据集进行特征降维,模型的最后使用了三种分类器进行连接,该模型使用降维后的训练集训练分类器,最后用测试集进行预测。实验结果表明,该方法在维数较大、特征信息冗余的数据集上的预测性能要优于基准的软件缺陷预测模型和基于现有的特征提取方法的软件缺陷预测模型,并且适用于不同分类算法。     

基于Markov随机场模型的纹理图像的缺陷检测

在工业自动化研究中,部件的缺陷检测是非常重要的过程。文中提出了一种基于图像纹理分析的表面缺陷检测方法,图像表面纹理特征是利用Markov随机场模型来描述的,通过学习和聚类分析来检测出纹理图像中有缺陷的区域。试验结果表明,该方法可以有效地描述不同种物质表面的纹理特征,并能准确地检测和定位缺陷。