Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index; 10 of the 14 sources its claims cite were checked for support.

Automatic Detection and Classification of Steel Surface Defect Using Deep Convolutional Neural Networks

Open access
Wang, Shuai, Xia, Xiaojun, Ye, Lanqing, Yang, Binbin · 2021 · Metals

Abstract

Automatic detection of steel surface defects is very important for product quality control in the steel industry. However, the traditional method cannot be well applied in the production line, because of its low accuracy and slow running speed. The current, popular algorithm (based on deep learning) also has the problem of low accuracy, and there is still a lot of room for improvement. This paper proposes a method combining improved ResNet50 and enhanced faster region convolutional neural networks (faster R-CNN) to reduce the average running time and improve the accuracy. Firstly, the image input into the improved ResNet50 model, which add the deformable revolution network (DCN) and improved cutout to classify the sample with defects and without defects. If the probability of having a defect is less than 0.3, the algorithm directly outputs the sample without defects. Otherwise, the samples are further input into the improved faster R-CNN, which adds spatial pyramid pooling (SPP), enhanced feature pyramid networks (FPN), and matrix NMS. The final output is the location and classification of the defect in the sample or without defect in the sample. By analyzing the data set obtained in the real factory environment, the accuracy of this method can reach 98.2%. At the same time, the average running time is faster than other models.

Overview

CoChat Summary

TL;DR

The paper proposes a two-stage pipeline combining an improved ResNet50 classifier with an enhanced Faster R-CNN detector for steel surface defects, using a 0.3 defect-probability threshold to skip detection on likely defect-free samples. On the Severstal Kaggle dataset, the combined model reaches 0.982 accuracy with 63.3 ms average running time.

Key findings

  • The combined classification + object detection model achieves 0.982 accuracy and 63.3 ms average running time, compared with 0.9752 accuracy / 2.9 ms for the improved ResNet50 alone, 0.972 accuracy / 214.5 ms for improved Faster R-CNN alone, and 0.97 accuracy / 209.4 ms for Wei et al.
  • The classification model ResNet_vd_dcnV2_ImprovedCutout reaches 0.9752 accuracy with test time augmentation (TTA), with TTA increasing accuracy by about 0.05%; its running time is 2.9 ms.
  • The final object detection model reaches mAP 0.876, with AP values of 0.845 for pitted surface, 0.829 for crazing, 0.913 for scratches, and 0.916 for patches; YOLOv3 scores mAP 0.704 and Faster-RCNN 0.827.
  • In the combined pipeline, 7224 samples are predicted defect-free with accuracy 0.991 (7159 correct, 65 wrong); the remaining 2830 samples are processed by detection with accuracy 0.961, yielding final whole-model accuracy 0.982.

Methods

  • Dataset: Kaggle “Severstal: Steel Defect Detection,” 12,568 grayscale images of size 1600 × 256, split into four 400 × 256 images, giving 50,272 samples: 37,080 without defects, 12,876 with one defect type, and 316 with two defect types; defect classes include pitted surface, crazing, scratches/scrapes, and patches; data split 80% training / 20% validation.
  • Pipeline: images are resized to 224 × 224 and classified by improved ResNet50 with deformable convolution network (DCN) and improved cutout; if defect probability is less than 0.3, the sample is output as defect-free; otherwise it is resized to 800 × 1344 and passed to improved Faster R-CNN with spatial pyramid pooling (SPP), enhanced feature pyramid networks (FPN), matrix NMS, CoordConv, and modified anchors.
  • Training: transfer learning from ImageNet; classification model trained for 80 epochs with learning rate 0.01 and Momentum 0.9; object detection model trained for 50 epochs with learning rate 0.0025, batch_size 4, and warmup 2500; data augmentation and weighted cross entropy / focal loss are used to address class imbalance.

Limitations

  • The paper states dataset annotation problems: some pitted surface defects are not marked, some areas marked with defects have problems, and some images with defects are placed in the no-defect samples; future work will further improve the dataset and study model compression and acceleration.
  • The paper states pitted surface defects are usually small and not obvious, so accuracy for this class is not high; cracking defects are narrow and long, also limiting accuracy; the final detection accuracy is lower than using the detection model on all samples because the remaining no-defect samples are more difficult to detect.
  • Evident from the text: evaluation uses a single public Kaggle dataset rather than a production-line dataset, so generalization to real factory conditions is not demonstrated; class imbalance also remains substantial, with crazing defects accounting for only 1.6% of defect samples before augmentation to 800.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/met11030388

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

8 Valid · 4 Not valid · 2 Not checked · 50 Own results and interpretations

36 more claims

Create a free account to see all 64 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 20
  • Probably not a substantial claim. 2
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 44 references

Cited by

See all 184

Metadata

Authors
Wang, S., Xia, X., Ye, L., & Yang, B.
Affiliations
Chinese Academy of Sciences (CN)
Corresponding author
Wang, Shuai
Venue
Metals
Publisher
MDPI AG
Published
2021-02-26
Volume
11
Issue
3
Pages
388
ISSN
2075-4701 (ISSN-L)
OpenAlex
W3134639409
Cited by
184
References
44
Citations by year
2026: 16 · 2025: 34 · 2024: 58 · 2023: 39 · 2022: 25 · 2021: 12
FWCI
17.29
Type
journal article
Language
en
Open access
Locations
  • Metals (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Department of science & technology of Liaoning Province, China (XLYC180839)
Topics
  • Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Engineering › Mechanical Engineering › Non-Destructive Testing Techniques
Keywords
steel surface defect detection, defect classification, ResNet50, Faster R-CNN, deformable convolutional networks, spatial pyramid pooling, feature pyramid network, non-maximum suppression, steel quality control, industrial production lines

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Automatic Detection and Classification of Steel Surface Defect Using Deep Convolutional Neural Networks. https://app.cochat.ai/papers/10.3390/met11030388
The paper
Wang, S., Xia, X., Ye, L., & Yang, B. (2021). Automatic Detection and Classification of Steel Surface Defect Using Deep Convolutional Neural Networks. Metals. https://doi.org/10.3390/met11030388

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.