In recent years, a multitude of self-supervised anomaly detection algorithms have been proposed. Among them, PatchCore has emerged as one of the state-of-the-art methods on the widely used MVTec AD benchmark due to its efficient detection capabilities and cost-saving advantages in terms of labeled data. However, we have identified that the PatchCore similarity principal approach faces significant limitations in accurately locating anomalies when there are positional relationships between similar samples, such as rotation, flipping, or misaligned pixels. In real-world industrial scenarios, it is common for samples of the same class to be found in different positions. To address this challenge comprehensively, we introduce Feature-Level Registration PatchCore (FR-PatchCore), which serves as an extension of the PatchCore method. FR-PatchCore constructs a feature matrix that is extracted into the memory bank and continually updated using the optimal negative cosine similarity loss. Extensive evaluations conducted on the MVTec AD benchmark demonstrate that FR-PatchCore achieves an impressive image-level anomaly detection AUROC score of up to 98.81%. Additionally, we propose a novel method for computing the mask threshold that enables the model to scientifically determine the optimal threshold and accurately partition anomalous masks. Our results highlight not only the high generalizability but also substantial potential for industrial anomaly detection offered by FR-PatchCore.
This paper proposed a novel anomaly detection (AD) approach of high-speed train images based on convolutional neural networks and the Vision Transformer. Different from previous AD works, in which anomalies are identified with a single image using classification, segmentation, or object detection methods, the proposed method detects abnormal difference between two images taken at different times of the same region. In other words, we cast anomaly detection problem with a single image into a difference detection problem with two images. The core idea of the proposed method is that the “anomaly” commonly represents an abnormal state instead of a specific object, and this state should be identified by a pair of images. In addition, we introduced a deep feature difference AD network (AnoDFDNet) which sufficiently explored the potential of the Vision Transformer and convolutional neural networks. To verify the effectiveness of the proposed AnoDFDNet, we gathered three datasets, a difference dataset (Diff dataset), a foreign body dataset (FB dataset), and an oil leakage dataset (OL dataset). Experimental results on the above datasets demonstrate the superiority of the proposed method. In terms of the F1-score, the AnoDFDNet obtained 76.24%, 81.04%, and 83.92% on Diff dataset, FB dataset, and OL dataset, respectively.
In high‐speed train safety inspection, two changed images which are derived from corresponding parts of the same train and photographed at different times are needed to identify whether they are defects. The critical challenge of this change classification task is how to make a correct decision by using bitemporal images. In this paper, two convolutional neural networks are presented to perform this task. Distinct from traditional classification tasks which simply group each image into different categories, the two presented networks are capable of inherently detecting differences between two images and further identifying changes by using a pair of images. In doing so, even in the case that abnormal samples of specific components are unavailable in training, our networks remain capable to make inference as to whether they become abnormal using change information. This proposed method can be used for recognition or verification applications where decisions cannot be made with only one image (state). Equipped with deep learning, this method can address many challenging tasks of high‐speed train safety inspection, in which conventional methods cannot work well. To further improve performance, a novel multishape training method is introduced. Extensive experiments demonstrate that the proposed methods perform well.
10 August 2022 in Neural Computing and Applications
Moire profilometry (MP) is one of the three-dimensional (3D) topography measurement methods of structured light, which has the advantages of single frame reconstruction, high speed, no contact and high precision, and is suitable for dynamic measurement scenes. In this article, the digital MP is applied to the wheel tread measurement, the virtual grating is generated by computer to project to the object surface, the moire fringe pattern of the object is obtained by filtering, and finally the continuous phase pattern is obtained by phase unwrapping. The 3D shape reconstruction of the wheel tread is realized, and a new method of wheel tread detection is provided. At the same time, in this paper, the results of using different filters are compared, and the significance of the frequency domain filtering to MP is proved. It is necessary to choose a suitable filtering method according to different environmental conditions. At present, digital MP can be used in industrial static detection, and it can be extended to the dynamic detection of rolling wheels in the future, so as to improve the detection efficiency and realize the automatic detection of trains.
Phase shifting profilometry (PSP) is widely used in three-dimensional (3D) optical metrology applications such as mechanical engineering, industrial monitoring, computer vision, and biomedicine because of its high measurement accuracy. In PSP, multi-frequency temporal phase unwrapping (TPU) plays a dominant role due to its high-accuracy measurement of surfaces with discontinuities and isolated objects. However, it requires a large number of fringe patterns. To reduce the number of required patterns, a new dual-frequency composite-pattern TPU method is developed, which needs only three patterns. The new method combines two fringe patterns of different frequencies into three composite fringe patterns, reconstructs rough 3D points by demodulating the low-frequency fringe pattern using geometric constraints, and reprojects it to obtain a rough low-frequency absolute phase, which is then refined to correct edge errors. Finally, the high-frequency wrapped phase is unwrapped under the guidance of the refined low-frequency phase, thereby the accurate 3D points are reconstructed based on stereo-vision technology. Experimental results demonstrated that the proposed method can achieve 3D surface measurement with only three images, which greatly improves the measurement efficiency. The proposed method has great application potential for the rapid 3D measurement of discontinuous surfaces and isolated objects.