The rapid growth of railway operation mileage in recent years, due to the construction of numerous new railway lines in many countries, has significantly increased the pressure on maintenance. During the daily operation of railway systems, the interaction between wheels and rails inevitably leads to surface defects such as spalling, corrugation, and grinding, which pose serious hidden dangers to safe operation. Unlike internal defects in rails that can be detected using techniques such as ultrasound [1] and eddy current ^[2][3][2,3], traditional rail surface defect detection is mainly conducted through manual visual inspection, which is inefficient and heavily relies on human workers’ experience [4]. In recent years, many researchers have focused on developing machine vision-based rail surface defect detection technologies that offer higher efficiency and accuracy to address the aforementioned issues. With the rapid development of artificial intelligence technology, deep-learning-based algorithms, specifically supervised learning-based defect detection algorithms, are being widely applied in rail surface defect detection ^[5][6][7][8][5,6,7,8].

However, defect samples are difficult to obtain in practical work; thus, defect detection methods based on supervised learning face two important challenges due to insufficient defect samples. One of them is the risk of overfitting caused by the limited training data, which may not adequately represent the distribution of defect; additionally, supervised learning methods typically require the use of a portion of the defect data for training, leading to a reduction in the number of testing samples available for validation, which affects the credibility of the validation results. Inspired by the concept of anomaly detection (AD), some researchers have turned their attention to utilizing unsupervised learning techniques to address the aforementioned issues in the field of defect detection ^[9][10][11][9,10,11]. However, these unsupervised learning-based methods rely completely on modeling the distribution of normal samples, lacking an understanding of defect data, which may lead to poor classification performance and a potentially high false-positive/negative rate [12].

Previous research on rail surface defect detection often utilizes traditional image processing techniques to extract features from defect images and trains detection models using corresponding machine learning methods ^{[13][14][15][16]}[13,14,15,16]. However, the performance of these methods is limited by the design of feature extraction, and the detection results can easily be affected by factors such as lighting, noise, and other factors. With the rapid development of deep learning technology, an increasing number of researchers have started studying rail surface defect detection methods based on deep learning, especially supervised learning methods. Wang Hao et al. integrated the improved pyramid feature fusion and modified loss function into the Mask-RCNN algorithm for the purpose of detecting rail surface defects [4]. Meng Si et al. proposed a multi-task architecture for rail surface defect detection, which includes two branch models for rail detection and defect segmentation [17]. Zhang Hui et al. cascaded the one-stage object detection algorithms SSD and YOLOv3, integrating the detection results from both networks to improve the accuracy of rail surface defect detection [18]. However, these approaches neglected the fact that defect samples are scarce and difficult to obtain in practical work. Due to the limited number of defect samples in the field of defect detection, supervised algorithms-based defect detection models often face issues of overfitting and low validation credibility. To address these problems, many researchers have proposed corresponding solutions. D. Zhang et al. partitioned the rail image data into multiple segments and trained the defect detection model. However, this approach did not fundamentally solve the problem [19], and more researchers have recently started studying steel rail surface defect algorithms based on unsupervised anomaly detection algorithms. Q. Zhang et al. implemented the detection of rail surface defects using the multi-scale cross FastFlow model [20], while Menghui Niu et al. proposed an unsupervised stereoscopic saliency detection method for detecting rail surface defects and achieved good detection results [21]. However, some studies have pointed out that unsupervised anomaly defect detection algorithms often lead to a higher false detection rate ^[22][23][22,23] due to the lack of knowledge about defect samples during the training process.

Unsupervised Anomaly Detection for Industrial Images

Deep-learning-based algorithms are being widely used in industrial defect detection research in recent years due to their high efficiency and accuracy. Many researchers have devoted themselves to researching industrial defect detection algorithms based on supervised learning algorithms, which significantly depends on labeled defect data ^{[24][25][26][27][28][29]}[24,25,26,27,28,29]. However, due to the hardship of collecting defective samples, it is extremely hard to obtain enough defect data for a deep model to learn its distribution. Furthermore, supervised learning-based methods require defect data for training, which further restricts the quantity of test datasets and affects the credibility of validation performance. In recent years, unsupervised-based anomaly detection (AD) algorithms have become the mainstream paradigm for industrial defect detection, which can be categorized as reconstruction-based and feature-embedding-based ^[30][31][32][30,31,32]. Reconstruction-based methods aim to train a deep network such as an adversarial generative network (GAN) or auto encoder (AE) to reconstruct normal images. When defective images are fed into the network, the defective parts cannot be reconstructed well, allowing for the detection of defects. However, sometimes the model can also yield a good reconstruction for the defective parts due to the powerful ability of the deep model [30]. Feature-embedding-based methods became the prevalent architecture in recent years, which typically consisted of a feature extractor and a feature estimator. A feature extractor is a deep network, typically a ResNet [33], that is pre-trained on ImageNet datasets. It is used to extract features from normal images, which are then stored into a memory bank. A feature estimator is used to estimate the distribution for normal features, which can be a multidimensional Gaussian distribution [34], clustering methods [35], or flow-based methods [36]. To avoid the deviation caused by different data distribution between industrial images and ImageNet datasets, only features from shallow layers are used. After distribution estimation, a distance metric is typically used to detect defects, since defects should be far from the center of the estimated distribution. One major drawback of embedding-based anomaly detection algorithms is that they estimate the distribution separately for each patch of the feature map, resulting in a massive and redundant feature memory bank to restore features from each patch. Many researchers have tried different methods to alleviate the problem: Padim experimentally studied the possibility to reduce redundancy of the memory bank and eventually chose to randomly discard a portion of the extracted features [30]; Patchcore utilized a coreset subsampling method to select representative features [32], thereby compressing the size of the feature memory bank. The paper introduces a feature representation method widely used in few-shot learning, which obtains a representative and compact feature memory bank and alleviates the aforementioned redundancy problem of the memory bank for rail surface defect detection.

Few-Shot Learning

In recent years, deep learning algorithms based on supervised learning have garnered significant attention from researchers due to the remarkable ability of deep models and large-scale datasets with high-quality labels. However, it is well known that supervised algorithms fail to acquire strong generalization ability when trained on a dataset with a small amount of data. Moreover, in many fields such as industrial defect detection, collecting a large-scale dataset with high-quality annotations proves to be challenging. This realization has prompted many researchers to shift their focus to the field of few-shot learning, with the aim of enabling the model to obtain strong generalization ability with only a few samples, akin to human beings. Within the domain of few-shot learning in computer vision, image classification tasks are a prominent focal point. These tasks can be broadly categorized into three distinct classes: data-augmentation-based methods, parameter-optimization-based methods, and metric-learning-based methods. Data-augmentation-based methods aim to address the challenge of limited samples in few-shot learning indirectly by enhancing the intricacy of the dataset through data augmentation. Trinet [37] employs autoencoders to map the features to the semantic space, followed by mapping the augmented features back to the sample space via semantic nearest neighbor search. Moreover, Patchmix [38] resolves the issue of distribution shift by substituting a specific region of the query image with random gallery images from diverse categories. Parameter-optimization-based methods generally first train a meta-learner to learn common features (prior knowledge) of different tasks and then apply the obtained meta-knowledge to fine-tune the base learner on the query set. The model-agnostic meta-learning (MAML) [39], which first trains the model on a large number of task sets to obtain an adaptable weight and then fine-tunes the model on the target task to obtain the final classifier. Metric-learning-based methods leverage pre-trained neural networks to extract features from training data. These extracted features are then utilized to measure similarity between the training data and test data using a metric. Representative methods include Siamese networks [40] and matching networks [41]. The former inputs two samples into the neural network and compares the similarity of the output feature vectors, while the latter uses attention mechanisms to obtain information about the correlation between feature vectors. A typical embedding-based approach to few-shot image classification is the prototypical network [42], which utilizes a pre-trained model to extract features from a limited amount of labeled data and learns corresponding feature prototypes from them. The network then produces a distribution over classes for an input feature based on a softmax function over distances to the prototypes in the embedding space. The prototypical network approach, combined with the utilization of mask average pooling, has been widely adopted in few-shot semantic segmentation methods. In addition, the idea of prototype features in prototypical networks has also been widely applied in many unsupervised anomaly detection algorithms ^[43][44][43,44].