Apnet: Generating Precise Anomaly Prior Information for Mixed-Supervised Defect Detection

The Metropolis Monte Carlo algorithm was applied to produce tomographic reconstructions from scarce projection data supplemented by prior information about the smoothness of the object. The prior information is represented by means of local energy functions, which are added to the projection error. The proposed prior function is an extension of previous proposals of border filters, the novelty introduced here being an adaptive control of the filter during the reconstruction process. The method was tested on synthetic phantoms and the reconstructions of a real object from a small number of projections. The technique shows good results in images with piecewise homogeneous regions, and can be useful in certain applications, where the scanning views are within an angular range that is either limited or sparsely sampled, as the detection of material defects in non-destructive testing or special anatomical components in medical images. Finally, the method is applied to the reconstruction of an industrial application of a stainless-steel BNC elbow from very few projections.

Subsea pipelines can be inspected via 2D cross-sectional x-ray computed tomography (CT). Traditional reconstruction methods produce an image of the pipe’s interior that can be post-processed for detection of possible defects. In this paper we propose a novel Bayesian CT reconstruction method with built-in defect detection. We decompose the reconstruction into a sum of two images; one containing the overall pipe structure, and one containing defects, and infer the images simultaneously in a Gibbs scheme. Our method requires that prior information about the two images is very distinct, i.e. the first image should contain the large-scale and layered pipe structure, and the second image should contain small, coherent defects. We demonstrate our methodology with numerical experiments using synthetic and real CT data from scans of subsea pipes in cases with full and limited data. Experiments demonstrate the effectiveness of the proposed method in various data settings, with reconstruction quality comparable to existing techniques, while also providing defect detection with uncertainty quantification.

Cheese yarn is made up of cotton yarn by winding method which makes it texture element coarse, which leads to trouble in the defects detection process. Therefore, a novel cheese yarn surface defects detection algorithm based on multidirectional matching filter (MMF) was proposed. Firstly, the source image is processed by OTSU threshold method, and the ellipse circle fitting method is used to determine the position of the ring-shaped surface of cheese yarn. Then, polar coordinate transformation is used to expand the ring-shaped surface to a rectangle. According to the texture characteristics of cheese yarn, different-of-Gaussian (DoG) operator is used to extract the edge image. After that, the angle range of 0 to 180 is quantified into several directional template images which are used to correlate with edge image, and the maximum correlation result of each pixel is elected to the maximum correlation image. The defects are segmented out from the maximum correlation image by an empirical threshold according to prior information. Finally, using the texture of the cheese yarn and prior information to refine the multidirectional matching filter image, using the shape information of connected domain to determine the location of the defects. The experiment shows that our method can effectively detect the “net-yarn” surface defects caused by improper forming process of the cheese yarn. The algorithm proposed in this paper can detect the surface defects of cheese with high realtime performance, which is suitable for on-line detection of cheese defects.

The problem of image reconstruction in SPECT is considered in the general context of statistical regularization of ill-posed problems. The MAP algorithm with the prior information based on the maximum entropy (MEENT) concept is presented in this work. The MENT method is known as a powerful tool for solving tomography problems. Introduction of prior information entails the necessity of choosing an appropriate regularization parameter. The effectiveness of a reconstruction method depends strongly on the choice of a good parameter. In practice optimal regularization parameters are often found empirically: one needs to test a large number of parameters in order to find a reasonable good one. This paper intends to make some further contribution to the subject in developing some practical regularization parameter choice strategies in SPECT. The method for selecting an optimal regularization parameter based on the theory of a 'laminar ensemble' is studied. This statistical method has an excellent theoretical basis. Numerical tests have shown, that there is a tradeoff between the resolution and noise level of the image that changes with iteration. The regularization parameter enables us to operate the tradeoff. This nice property seems especially useful and important in real applications as it provides a feasible and stable numerical resolution. The optimal regularization parameter was defined at each iteration step automatically without requiring the user to select it. The adaptive chi-square criterion was used to control and stop the iteration process. The MLAP-MENT algorithm has been tested using the 3D heart phantom. The important aim of the work was to evaluate the potential of the algorithm in the defect detection. The defect was modeled as a small region of decreased count density. In numerical experiments we investigated the convergence properties of the algorithm. A comparison to the maximum-likelihood-based (OS EM) algorithm was performed.

The vision-based defect detection of textile surface is an essential problem in evaluating the appearance quality. In previous studies on vision-based defect detection, the following two points were neglected: (1) The proportion of defects on fabric is small, resulting in a low signal-to-noise ratio of defect images and a lack of sampling features; (2) the irregular shape of defects can overlap in position, leading to localization difficulty. In this paper, we propose a prior knowledge-embedded deformable convolutional network (PKE-DCNet) based on deep learning to address these two issues. First, a feature extraction method with prior information is designed with a shape-matching deep clustering module and a region-biased sampling module for detecting defects with complex shape. Then, a defect boundary boxes adaptive generation method is proposed with an anchor-free search mechanism and irregular edge contour computation to detect the key points in the region. Extensive experiments with mixed fabric defect datasets demonstrated that PKE-DCNet reached an overall mAP of 95.36% for six types of defects within a detection speed of 322 FPS, which was better than state-of-the-art methods.

ABSTRACTInspections and testing are two of the most commonly performed software quality assurance processes today. Typically, these processes are applied in isolation, which, however, fails to exploit the benefits of systematically combining and integrating them. In consequence, tests are not focused on the basis of early defect detection data. Expected benefits of such process integration include higher defect detection rates or reduced quality assurance effort. Moreover, when conducting testing without any prior information regarding the system's quality, it is often unclear how to focus testing. A systematic integration of inspection and testing processes requires context‐specific knowledge about the relationships between inspections and testing. This knowledge is typically not available and needs to be empirically identified and validated. Often, context‐specific assumptions can be seen as a starting point for generating such knowledge. On the basis of the integrated inspection and testing approach, In2Test, which uses inspection data to focus testing, we present in this article how knowledge about the relationship between inspections and testing can be gained, documented, and evolved in an analytical or empirical manner. In addition, this article gives an overview of related work and highlights future research directions. Copyright © 2012 John Wiley & Sons, Ltd.

To achieve enhanced accuracy of fabric representation and defect detection, an innovative approach using a sparse dictionary with small patches was used for fabric texture characterisation. The effectiveness of the algorithm proposed was tested through comprehensive characterisation by studying eight weave patterns: plain, twill, weft satin, warp satin, basket, honeycomb, compound twill, and diamond twill and detecting fabric defects. Firstly, the main parameters such as dictionary size, patch size, and cardinality T were optimised, and then 40 defect-free fabric samples were characterised by the algorithm proposed. Subsequently, the impact of the weave pattern was investigated based on the representation result and texture structure. Finally, defective fabrics were detected. The algorithm proposed is an alternative simple and scalable method to characterise fabric texture and detect textile defects in a single step without extracting features or prior information.

Automatic recognition of defects behind railway tunnel linings in GPR images using transfer learning

Bayesian hierarchical hyper-Laplacian priors for high-resolution defect imaging in pipe structures

To solve the problem of anomaly detection in annular metal turning surfaces, this paper develops an anomaly detection algorithm based on a priori information and a multi-scale self-referencing template by combining the imaging characteristics of annular workpieces. First, the annular metal turning surface is unfolded into a rectangular expanded image using bilinear interpolation to facilitate subsequent algorithm development. Second, the grayscale information from the positive samples is used to obtain the a priori information, and a multi-scale self-referencing template method is used to obtain its own multi-scale information. Then, the phase error and large-size anomaly interference problems of the self-referencing method are overcome by combining the a priori information with its own information, and an accurate response to anomalous regions of various sizes is realized. Finally, the segmentation completeness of the anomalous region is improved by utilizing the region growing method. The experimental results show that the proposed method achieves a mean pixel AUROC of 0.977, and the mean M_IOU of segmentation reaches 0.788. In terms of efficiency, this method is also much more efficient than the commonly used anomaly detection algorithms. The proposed method can achieve rapid and accurate detection of defects in annular metal turning surfaces and has good industrial application value.

Structured illumination using sinusoidal patterns has been used for optical imaging of biological tissues in biomedical research, and of horticultural products in food quality evaluation. Implementation of structured-illumination imaging relies on retrieval of amplitude images, which is conventionally achieved by a phase-shifting technique that requires collecting a minimum of three phase-shifted images. In this study, we have proposed Gram-Schmidt orthonormalization (GSO) to retrieve amplitude component (AC) images using only two phase-shifted images. We have proposed two forms of GSO implementation, and prior to GSO processing, we eliminated the direct component (DC) background by subtracting a DC image we recovered using a spiral phase function (SPF) in the Fourier space. We demonstrated the GSO methods through numerical simulations and application examples of detection of bruise defects in apples by structured-illumination reflectance imaging (SIRI). GSO performed comparably to conventional three-phase-based demodulation. It is simple, fast and effective for amplitude retrieval and requires no prior phase information, which could facilitate fast implementation of structured-illumination imaging.

Wood has long been an important building material. The wide variety in appearance, strength properties and the possibility of different dimensions are some of the reasons why wood is still very attractive for this purpose. Computed tomography offers great potential for non-destructive testing of the internal structure of wood. In the paper, we apply the computed tomography (CT) technology for log nondestructive testing and calculate the CT number range of different wood tomography slices in statistic method, then the parameters of CT window level and window width in wood testing procedure were set. Because of the linear relationship between wood density and CT number the fitting linear formula between CT number and wood density was calculated. As for the defects in the wood X-ray computed tomography scanning technology is applied to the detection of internal defects for the purpose of obtaining prior information that can be used to arrive at better log sawing decision. A method in wood CT image edge detection of defects based on multifractal theory is applied in the paper.

Wood nondestructive testing technology is a new and comprehensive subject. X-ray computed tomography (CT) scanning technology has been applied to the detection of internal defects in the logs for the purpose of obtaining prior information that can be used to arrive at better log sawing decision. Thus, the recognition of internal defects becomes more and more important work. A method in log CT image edge detection of defects based on multifractal theory was applied in the paper. The Holder exponent of image pixels was computed first, then its multifractal spectrum was estimated and different image pixels were classified, Based on multifractal theory, the set of both singular edge points and smoothing edge points is the set of image edge points. Experimental result showed that the method of log CT image in the edge detection based on multifractal theory was a more effective and more local method.

A novel framework for low-contrast and random multi-scale blade casting defect detection by an adaptive global dynamic detection transformer

With the growing demand for efficient and high-precision quality control in textiles, the limitations of traditional fabric defect detection methods have become increasingly apparent, particularly in handling complex backgrounds, diverse defect types, and small-object detection. To address these challenges, this study proposes a fabric defect detection method based on morphological prior features. By deeply integrating morphological prior information into the feature extraction and fusion modules, the robustness and accuracy of defect detection are improved significantly. Specifically, the divergent path feature enhancement module enhances the detection of small-object defects through fine-grained feature extraction. The oriented space and multiscale feature extraction module combines multiscale and directional feature extraction techniques to improve the recognition of defects with extreme aspect ratios. Furthermore, the efficient multipath feature fusion network achieves comprehensive capture of defect features by integrating shallow and deep features. In addition, the proposed fabric mosaic data augmentation strategy dynamically adjusts the cropping offset points, effectively preserving the pixel and feature integrity of target defects under conditions of data scarcity and imbalance. Experimental results demonstrate that the proposed model achieves a precision of 75.5%, a recall of 75.8%, and an F1 score of 75.6%. The mAP@.5 is improved by 9.8% compared with the baseline model. The proposed approach strikes a favorable balance between detection accuracy, computational complexity, and inference speed, showcasing excellent generalization capability and practical application potential.