Surface Defect Detection Methods for Industrial Products: A Review

Detection Method of Stator Coating Quality of Flat Wire Motor Based on Improved YOLOv8s

The stator of a flat wire motor is the core component of new energy vehicles. However, detecting quality defects in the coating process in real-time is a challenge. Moreover, the number of defects is large, and the pixels of a single defect are very few, which make it difficult to distinguish the defect features and make accurate detection more difficult. To solve this problem, this article proposes the YOLOv8s-DFJA network. The network is based on YOLOv8s, which uses DSFI-HEAD to replace the original detection head, realizing task alignment. It enhances joint features between the classification task and localization task and improves the ability of network detection. The LEFG module replaces the C2f module in the backbone of the YOLOv8s network that reduces the redundant parameters brought by the traditional BottleNeck structure. It also enhances the feature extraction and gradient flow ability to achieve the lightweight of the network. For this research, we produced our own dataset of stator coating quality regarding flat wire motors. Data augmentation technology (Gaussian noise, adjusting brightness, etc.) enriches the dataset, to a certain extent, which improves the robustness and generalization ability of YOLOv8s-DFJA. The experimental results show that in the performance of YOLOv8s-DFJA compared with YOLOv8s, the mAP@.5 index increased by 6.4%, the precision index increased by 1.1%, the recall index increased by 8.1%, the FPS index increased by 9.8FPS/s, and the parameters decreased by 3 Mb. Therefore, YOLOv8s-DFJA can be better realize the fast and accurate detection of the stator coating quality of flat wire motors.

Intelligent Inspection Method and System of Plastic Gear Surface Defects Based on Adaptive Sample Weighting Deep Learning Model

After injection molding, plastic gears often exhibit surface defects, including those on end faces and tooth surfaces. These defects encompass a wide range of types and possess complex characteristics, which pose challenges for inspection. Current visual inspection systems for plastic gears suffer from limitations such as single-category defect inspection and low accuracy. There is an urgent industry need for a comprehensive and accurate method and system for inspecting defects on plastic gears, with improved inspection capability and higher accuracy. This paper presents an intelligent inspection algorithm network for plastic gear defects (PGD-net), which effectively captures subtle defect features at arbitrary locations on the surface compared to other models. An adaptive sample weighting method is proposed and integrated into an improved Focal-IoU loss function to address the issue of low inspection accuracy caused by imbalanced defect dataset distributions, thus enhancing the regression accuracy for difficult defect categories. CoordConv layers are incorporated into each inspection head to improve the model's generalization capability. Furthermore, a dataset of plastic gear surface defects comprising 16 types of defects is constructed, and our algorithm is trained and tested on this dataset. The PGD-net achieves a comprehensive mean average precision (mAP) value of 95.6% for the 16 defect types. Additionally, an online inspection system is developed based on the PGD-net algorithm, which can be integrated with plastic gear production lines to achieve online full inspection and automatic sorting of plastic gear defects. The entire system has been successfully applied in plastic gear production lines, conducting daily inspections of over 60,000 gears.

Temporal-Quality Ensemble Technique for Handling Image Blur in Packaging Defect Inspection

Despite achieving numerous successes with surface defect inspection based on deep learning, the industry still faces challenges in conducting packaging defect inspections that include critical information such as ingredient lists. In particular, while previous achievements primarily focus on defect inspection in high-quality images, they do not consider defect inspection in low-quality images such as those containing image blur. To address this issue, we proposed a noble inference technique named temporal-quality ensemble (TQE), which combines temporal and quality weights. Temporal weighting assigns weights to input images by considering the timing in relation to the observed image. Quality weight prioritizes high-quality images to ensure the inference process emphasizes clear and reliable input images. These two weights improve both the accuracy and reliability of the inference process of low-quality images. In addition, to experimentally evaluate the general applicability of TQE, we adopt widely used convolutional neural networks (CNNs) such as ResNet-34, EfficientNet, ECAEfficientNet, GoogLeNet, and ShuffleNetV2 as the backbone network. In conclusion, considering cases where at least one low-quality image is included, TQE has an F1-score approximately 17.64% to 22.41% higher than using single CNN models and about 1.86% to 2.06% higher than an average voting ensemble.

DG2GAN: improving defect recognition performance with generated defect image sample

This article aims to improve the deep-learning-based surface defect recognition. In actual manufacturing processes, there are issues such as data imbalance, insufficient diversity, and poor quality of augmented data in the collected image data for product defect recognition. A novel defect generation method with multiple loss functions, DG2GAN is presented in this paper. This method employs cycle consistency loss to generate defect images from a large number of defect-free images, overcoming the issue of imbalanced original training data. DJS optimized discriminator loss is introduced in the added discriminator to encourage the generation of diverse defect images. Furthermore, to maintain diversity in generated images while improving image quality, a new DG2 adversarial loss is proposed with the aim of generating high-quality and diverse images. The experiments demonstrated that DG2GAN produces defect images of higher quality and greater diversity compared with other advanced generation methods. Using the DG2GAN method to augment defect data in the CrackForest and MVTec datasets, the defect recognition accuracy increased from 86.9 to 94.6%, and the precision improved from 59.8 to 80.2%. The experimental results show that using the proposed defect generation method can obtain sample images with high quality and diversity and employ this method for data augmentation significantly enhances surface defect recognition technology.

Utilizing Deep Learning for Defect Inspection in Hand Tool Assembly

The integrity of product assembly in the precision assembly industry significantly influences the quality of the final products. During the assembly process, products may acquire assembly defects due to personnel oversight. A severe assembly defect could impair the product's normal function and potentially cause loss of life or property for the user. For workpiece defect inspection, there is limited discussion on the simultaneous detection of the primary kinds of assembly anomaly (missing parts, misplaced parts, foreign objects, and extra parts). However, these assembly anomalies account for most customer complaints in the traditional hand tool industry. This is because no equipment can comprehensively inspect major assembly defects, and inspections rely solely on professionals using simple tools and their own experience. Thus, this study proposes an automated visual inspection system to achieve defect inspection in hand tool assembly. This study samples the work-in-process from three assembly stations in the ratchet wrench assembly process; an investigation of 28 common assembly defect types is presented, covering the 4 kinds of assembly anomaly in the assembly operation; also, this study captures sample images of various assembly defects for the experiments. First, the captured images are filtered to eliminate surface reflection noise from the workpiece; then, a circular mask is given at the assembly position to extract the ROI area; next, the filtered ROI images are used to create a defect-type label set using manual annotation; after this, the R-CNN series network models are applied to object feature extraction and classification; finally, they are compared with other object detection models to identify which inspection model has the better performance. The experimental results show that, if each station uses the best model for defect inspection, it can effectively detect and classify defects. The average defect detection rate (1-β) of each station is 92.64%, the average misjudgment rate (α) is 6.68%, and the average correct classification rate (CR) is 88.03%.

Automatic PCB Sample Generation and Defect Detection Based on ControlNet and Swin Transformer

In order to improve the efficiency and accuracy of multitarget detection of soldering defects on surface-mounted components in Printed Circuit Board (PCB) fabrication, we propose a sample generation method using Stable Diffusion Model and ControlNet, as well as a defect detection method based on the Swin Transformer. The method consists of two stages: First, high-definition original images collected in industrial production and the corresponding prompts are input to Stable Diffusion Model and ControlNet for automatic generation of nonindependent samples. Subsequently, we integrate Swin Transformer as the backbone into the Cascade Mask R-CNN to improve the quality of defect features extracted from the samples for accurate detection box localization and segmentation. Instead of segmenting individual components on the PCB, the method inspects all components in the field of view simultaneously over a larger area. The experimental results demonstrate the effectiveness of our method in scaling up nonindependent sample datasets, thereby enabling the generation of high-quality datasets. The method accurately recognizes targets and detects defect types when performing multitarget inspection on printed circuit boards. The analysis against other models shows that our improved defect detection and segmentation method improves the Average Recall (AR) by 2.8% and the mean Average Precision (mAP) by 1.9%.

Autoencoder-Based Unsupervised Surface Defect Detection Using Two-Stage Training

Accurately detecting defects while reconstructing a high-quality normal background in surface defect detection using unsupervised methods remains a significant challenge. This study proposes an unsupervised method that effectively addresses this challenge by achieving both accurate defect detection and a high-quality normal background reconstruction without noise. We propose an adaptive weighted structural similarity (AW-SSIM) loss for focused feature learning. AW-SSIM improves structural similarity (SSIM) loss by assigning different weights to its sub-functions of luminance, contrast, and structure based on their relative importance for a specific training sample. Moreover, it dynamically adjusts the Gaussian window's standard deviation (σ) during loss calculation to balance noise reduction and detail preservation. An artificial defect generation algorithm (ADGA) is proposed to generate an artificial defect closely resembling real ones. We use a two-stage training strategy. In the first stage, the model trains only on normal samples using AW-SSIM loss, allowing it to learn robust representations of normal features. In the second stage of training, the weights obtained from the first stage are used to train the model on both normal and artificially defective training samples. Additionally, the second stage employs a combined learned Perceptual Image Patch Similarity (LPIPS) and AW-SSIM loss. The combined loss helps the model in achieving high-quality normal background reconstruction while maintaining accurate defect detection. Extensive experimental results demonstrate that our proposed method achieves a state-of-the-art defect detection accuracy. The proposed method achieved an average area under the receiver operating characteristic curve (AuROC) of 97.69% on six samples from the MVTec anomaly detection dataset.

Identification and classification of surface defects for digital twin models of the workpiece

Workpiece surface defect detection is an indispensable part of intelligent production. The surface information obtained by traditional 2D image detection has some limitations due to the influence of environmental light factors and part complexity. However, the digital twin model has the characteristics of high fidelity and scalability, and the digital twin surface can be obtained by a device with a scanning accuracy of 0.02mm to achieve the representation of the real surface of the workpiece. The surface defect detection system for digital twin models is proposed based on the improved YOLOv5 model in this paper. Firstly, the digital twin model of the workpiece is reconstructed by the point cloud data obtained by the scanning device, and the surface features with defects are captured. Subsequently, the training dataset is calibrated based on the defect surface, where the defect types include Inclusion, Perforation, pitting surface and Rolled-in scale. Finally, the improved YOLOv5 model with CBAM mechanism and BiFPN module was used to identify the surface defects of the digital twin model and compare it with the original YOLOv5 model and other common models. The results show that the improved YOLOv5 model can realize the identification and classification of surface defects. Compared with the original YOLOv5 model, the mAP value of the improved YOLOv5 model has increased by 0.2%, and the model has high precision. On the basis of the same data set, the improved YOLOv5 model has higher recognition accuracy than other models, improving 11.7%, 3.4%, 6.2%, 33.5%, respectively. As a result, this study provides a practical and systematic detection method for digital twin model surface during the intelligent production process, and realizes the rapid screening of the workpiece with defects.

Image Data-Centric Visual Feature Selection on Roll-to-Roll Slot-Die Coating Systems for Edge Wave Coating Defect Detection

Roll-to-roll (R2R) manufacturing depends on a system's capability to deposit high-quality coatings with precise thickness, width, and uniformity. Therefore, consistent maintenance requires the immediate and accurate detection of coating defects. This study proposes a primary color selection (PCS) method to detect edge defects in R2R systems. This method addresses challenges associated with training data demands, complexity, and defect adaptability through a vision data-centric approach, ensuring precise edge coating defect detection. Using color information, high accuracy was achieved while minimizing data capacity requirements and processing time. Precise edge detection was facilitated by accurately distinguishing coated and noncoated regions by selecting the primary color channel based on color variability. The PCS method achieved superior accuracy (95.8%), outperforming the traditional weighted sum method (78.3%). This method is suitable for real-time detection in manufacturing systems and mitigates edge coating defects, thus facilitating quality control and production optimization.

Surface defect detection of hot rolled steel based on multi-scale feature fusion and attention mechanism residual block

To improve the precision of defect categorization and localization in images, this paper proposes an approach for detecting surface defects in hot-rolled steel strips. The approach uses an improved YOLOv5 network model to overcome the issues of inadequate feature extraction capacity and suboptimal feature integration when identifying surface defects on steel strips. The proposed method achieves higher detection accuracy and localization precision, making it more competitive and applicable in real production. Firstly, the multi-scale feature fusion (MSF) strategy is utilized to fuse shallow and deep features effectively and enrich detailed information relevant to target defects. Secondly, the CSPLayer Res2Attention block (CRA block) residual module is introduced to reduce the loss of defect information during hierarchical transmission, thereby enhancing the extraction of fine-grained features and improving the perception of details and global features. Finally, the experimental results indicate that the mAP on the NEU-DET and GC10-DET datasets approaches 78.5% and 67.3%, respectively, which is 4.9% and 2.1% higher than that of the baseline. Meanwhile, it has higher precision and more precise localization capabilities than other methods. Furthermore, it also achieves 59.2% mAP on the APDDD dataset, indicating its potential for growth in further domains.

An Overview of Image Generation of Industrial Surface Defects

Intelligent defect detection technology combined with deep learning has gained widespread attention in recent years. However, the small number, and diverse and random nature, of defects on industrial surfaces pose a significant challenge to deep learning-based methods. Generating defect images can effectively solve this problem. This paper investigates and summarises traditional defect generation and deep learning-based methods. It analyses the various advantages and disadvantages of these methods and establishes a benchmark through classical adversarial networks and diffusion models. The performance of these methods in generating defect images is analysed through various indices. This paper discusses the existing methods, highlights the shortcomings and challenges in the field of defect image generation, and proposes future research directions. Finally, the paper concludes with a summary.

Adaptive visual detection of industrial product defects

Visual inspection of the appearance defects on industrial products has always been a research hotspot pursued by industry and academia. Due to the lack of samples in the industrial defect dataset and the serious class imbalance, deep learning technology cannot be directly applied to industrial defect visual inspection to meet the real application needs. Transfer learning is a good choice to deal with insufficient samples. However, cross-dataset bias is unavoidable during simple knowledge transfer. We noticed that the appearance defects of industrial products are similar, and most defects can be classified as stains or texture jumps, which provides a research basis for building a universal and adaptive industrial defect detection model. In this article, based on the idea of model-agnostic meta-learning (MAML), we propose an adaptive industrial defect detection model through learning from multiple known industrial defect datasets and then transfer it to the novel anomaly detection tasks. In addition, the Siamese network is used to extract differential features to minimize the influence of defect types on model generalization, and can also highlight defect features and improve model detection performance. At the same time, we add a coordinate attention mechanism to the model, which realizes the feature enhancement of the region of interest in terms of two coordinate dimensions. In the simulation experiments, we construct and publish a visual defect dataset of injection molded bottle cups, termed BC defects, which can complement existing industrial defect visual data benchmarks. Simulation results based on BC defects dataset and other public datasets have demonstrated the effectiveness of the proposed general visual detection model for industrial defects. The dataset and code are available at https://github.com/zhg-SZPT/MeDetection.

Cost-Sensitive YOLOv5 for Detecting Surface Defects of Industrial Products

Owing to the remarkable development of deep learning algorithms, defect detection techniques based on deep neural networks have been extensively applied in industrial production. Most existing surface defect detection models assign equal costs to the classification errors among different defect categories but do not strictly distinguish them. However, various errors can generate a great discrepancy in decision risk or classification costs and then produce a cost-sensitive issue that is crucial to the manufacturing process. To address this engineering challenge, we propose a novel supervised classification cost-sensitive learning method (SCCS) and apply it to improve YOLOv5 as CS-YOLOv5, where the classification loss function of object detection was reconstructed according to a new cost-sensitive learning criterion explained by a label-cost vector selection method. In this way, the classification risk information from a cost matrix is directly introduced into the detection model and fully exploited in training. As a result, the developed approach can make low-risk classification decisions for defect detection. It is applicable for direct cost-sensitive learning based on a cost matrix to implement detection tasks. Using two datasets of a painting surface and a hot-rolled steel strip surface, our CS-YOLOv5 model outperforms the original version with respect to cost under different positive classes, coefficients, and weight ratios, but also maintains effective detection performance measured by mAP and F1 scores.

Surface defect detection method for electronic panels based on double branching and decoupling head structure

During the production of electronic panels, surface defects will inevitably appear. How to quickly and accurately detect these defects is very important to improve product quality. However, some problems such as high cost and low accuracy are still prominent when existing manual detection and traditional techniques are used to solve such problems. Therefore, more and more computer vision techniques are proposed to solve such problems, but the current application of deep learning-based object detection networks for surface defect detection of electronic panels is in a gap. The analysis found that there are two main reasons for this phenomenon. On the one hand, the surface defects of electronic panels have their unique characteristics such as multi-scale and irregular shape, and the current object detection networks cannot effectively solve these problems. On the other hand, the regression and classification tasks coupled in the current computational mechanism of each network are commonly found to cause the problem of conflict between them, which makes it more difficult to adapt these network models to the detection tasks in this scenario. Based on this, we design a supervised object detection network for electronic panel surface defect detection scenario for the first time. The computational mechanism of this network includes a prediction box generation strategy based on the double branch structure and a detection head design strategy that decouples the regression task from the classification task. In addition, we validated the designed network and the proposed method on our own collected dataset of surface defects in electronic panels. The final results of the comparative and ablation experiments show that our proposed method achieves an average accuracy of 78.897% for 64 surface defect categories, proving that its application to electronic panel surface defect detection scenarios can achieve better results.

Surface defect detection method for electronic panels based on attention mechanism and dual detection heads

Automatic detection of surface defects in electronic panels is receiving increasing attention in the quality control of products. The surface defect detection of electronic panels is different from other target detection scenarios and is a meaningful and challenging problem. Its main manifestation is that surface defects of electronic panels usually exhibit extreme irregularity and small target characteristics, which bring great difficulties to the task of surface defect target detection including feature extraction and so on. The traditional methods can only detect a very small number of defect classes under specific detection conditions. And due to the weak robustness of these methods, they cannot be applied in real production scenarios on a large scale. Based on this, this paper applies the target detection technique under deep learning to the surface defect detection scenario of electronic panels for the first time. At the same time, in order to make the designed target detection network applicable to the electronic panel surface defect detection scenario and to enhance the interpretability of the designed target detection network in terms of computer mechanism. We design a deformable convolution module with a convolutional self-attentive module to learn the offset and a dual detection head incorporating the SE (Squeeze-and-Excitation) mechanism for the irregular characteristics of electronic panel surface defects and the small target characteristics, respectively. Finally, we carried out a series of experiments on our own electronic panel defect data set, including comparison with the most advanced target detection algorithms and a series of ablation experiments against our proposed method. The final experimental results prove that our method not only has better interpretability, but also has better metric performance, in which the map_0.5 metric reaches 78.257%, which is an increase of 13.506 percentage points over YOLOV5 and 33.457 percentage points higher than Retinanet. The results prove the proposed method is effective.

Anomaly Detection of GAN Industrial Image Based on Attention Feature Fusion

As life becomes richer day by day, the requirement for quality industrial products is becoming greater and greater. Therefore, image anomaly detection on industrial products is of significant importance and has become a research hotspot. Industrial manufacturers are also gradually intellectualizing how product parts may have flaws and defects, and that industrial product image anomalies have characteristics such as category diversity, sample scarcity, and the uncertainty of change; thus, a higher requirement for image anomaly detection has arisen. For this reason, we proposed a method of industrial image anomaly detection that applies a generative adversarial network based on attention feature fusion. For the purpose of capturing richer image channel features, we added attention feature fusion based on an encoder and decoder, and through skip-connection, this performs the feature fusion for the encode and decode vectors in the same dimension. During training, we used random cut-paste image augmentation, which improved the diversity of the datasets. We displayed the results of a wide experiment, which was based on the public industrial detection MVTec dataset. The experiment illustrated that the method we proposed has a higher level AUC and the overall result was increased by 4.1%. Finally, we realized the pixel level anomaly localization of the industrial dataset, which illustrates the feasibility and effectiveness of this method.

Comparison of Pre-Trained YOLO Models on Steel Surface Defects Detector Based on Transfer Learning with GPU-Based Embedded Devices

Steel is one of the most basic ingredients, which plays an important role in the machinery industry. However, the steel surface defects heavily affect its quality. The demand for surface defect detectors draws much attention from researchers all over the world. However, there are still some drawbacks, e.g., the dataset is limited accessible or small-scale public, and related works focus on developing models but do not deeply take into account real-time applications. In this paper, we investigate the feasibility of applying stage-of-the-art deep learning methods based on YOLO models as real-time steel surface defect detectors. Particularly, we compare the performance of YOLOv5, YOLOX, and YOLOv7 while training them with a small-scale open-source NEU-DET dataset on GPU RTX 2080. From the experiment results, YOLOX-s achieves the best accuracy of 89.6% mAP on the NEU-DET dataset. Then, we deploy the weights of trained YOLO models on Nvidia devices to evaluate their real-time performance. Our experiments devices consist of Nvidia Jetson Nano and Jetson Xavier AGX. We also apply some real-time optimization techniques (i.e., exporting to TensorRT, lowering the precision to FP16 or INT8 and reducing the input image size to 320 × 320) to reduce detection speed (fps), thus also reducing the mAP accuracy.

A Novel Method to Inspect 3D Ball Joint Socket Products Using 2D Convolutional Neural Network with Spatial and Channel Attention

Product defect inspections are extremely important for industrial manufacturing processes. It is necessary to develop a special inspection system for each industrial product due to their complexity and diversity. Even though high-precision 3D cameras are usually used to acquire data to inspect 3D objects, it is hard to use them in real-time defect inspection systems due to their high price and long processing time. To address these problems, we propose a product inspection system that uses five 2D cameras to capture all inspection parts of the product and a deep learning-based 2D convolutional neural network (CNN) with spatial and channel attention (SCA) mechanisms to efficiently inspect 3D ball joint socket products. Channel attention (CA) in our model detects the most relevant feature maps while spatial attention (SA) finds the most important regions in the extracted feature map of the target. To build the final SCA feature vector, we concatenated the learned feature vectors of CA and SA because they complement each other. Thus, our proposed CNN with SCA provides high inspection accuracy as well as it having the potential to detect small defects of the product. Our proposed model achieved 98% classification accuracy in the experiments and proved its efficiency on product inspection in real-time.

Classification and Fast Few-Shot Learning of Steel Surface Defects with Randomized Network

Quality inspection is inevitable in the steel industry so there are already benchmark datasets for the visual inspection of steel surface defects. In our work, we show, contrary to previous recent articles, that a generic state-of-art deep neural network is capable of almost-perfect classification of defects of two popular benchmark datasets. However, in real-life applications new types of errors can always appear, thus incremental learning, based on very few example shots, is challenging. In our article, we address the problems of the low number of available shots of new classes, the catastrophic forgetting of known information when tuning for new artifacts, and the long training time required for re-training or fine-tuning existing models. In the proposed new architecture we combine EfficientNet deep neural networks with randomized classifiers to aim for an efficient solution for these demanding problems. The classification outperforms all other known approaches, with an accuracy 100% or almost 100%, on the two datasets with the off-the-shelf network. The proposed few-shot learning approach shows considerably higher accuracy at a low number of shots than the different methods under testing, while its speed is significantly (at least 10 times) higher than its competitors. According to these results, the classification and few-shot learning of steel surface defects can be solved more efficiently than was possible before.

Small object detection method with shallow feature fusion network for chip surface defect detection

The development of intelligent manufacturing often focuses on production flexibility, customization and quality control, which are crucial for chip manufacturing. Specifically, defect detection and classification are important for manufacturing processes in the semiconductor and electronics industries. The intelligent detection methods of chip defects are still challenge and have always been a particular concern of chip processing manufactures in an automated industrial production line. YOLOv4 method has been widely used for object detection due to its accuracy and speed. However, there are still difficulties and challenges in the detection for small targets, especially defects on chip surface. This study proposed a small object detection method based on YOLOv4 for small object in order to improve the performance of detection. It includes expanding feature fusion of shallow features; using k-means++ clustering to optimize the number and size of anchor box; and removing redundant YOLO head network branches to increase detection efficiency. The results of experiments reflect that SO-YOLO is superior to the original YOLOv4, YOLOv5s, and YOLOv5l models in terms of the number of parameters, classification and detection accuracy.

A Novel Learning Based Non-Lambertian Photometric Stereo Method for Pixel-Level Normal Reconstruction of Polished Surfaces

High-quality reconstruction of polished surfaces is a promising yet challenging task in the industrial field. Due to its extreme reflective properties, state-of-the-art methods have not achieved a satisfying trade-off between retaining texture and removing the effects of specular outliers. In this paper, we propose a learning based pixel-level photometric stereo method to estimate the surface normal. A feature fusion convolutional neural network is used to extract the features from the normal map solved by the least square method and from the original images respectively, and combine them to regress the normal map. The proposed network outperforms the state-of-the-art methods on the DiLiGenT benchmark dataset. Meanwhile, we use the polished rail welding surface to verify the generalization of our method. To fit the complex geometry of the rails, we design a flexible photometric stereo information collection hardware with multi-angle lights and multi-view cameras, which can collect the light and shade information of the rail surface for photometric stereo. The experimental results indicate that the proposed method is able to reconstruct the normal of the polished surface at the pixel level with abundant texture information.