PaperlayerView exact revision on arXiv

Source-linked AI summary

TEEP-RCNN: Texture-Enhanced Edge-aware Perception for Steel Surface Defect Detection via Improved Convolutional Block Attention in Faster R-CNN

Kirtan Rajesh

arXiv:2609.28077v1cs.CV

TL;DR

Steel defect detection remains challenging because defect textures differ and crazing is distributed across many locations. TEEP-RCNN modifies CBAM within a two-stage detector and combines differential learning-rate training with WBF-TTA. On NEU-DET, it reaches 73.3% mAP@50 in 10 epochs and exceeds YOLOv11m on rolled-in-scale mAP under the COCO metric.

  • Problem

    Standard two-stage detectors use generic feature representations across defect categories despite distinct texture patterns, including distributed crazing and localized scratches.

    Source: §I · pp. 1–2

  • Method

    TEEP-RCNN combines an improved CBAM with a differential-learning-rate training protocol and WBF-TTA inference in a Faster R-CNN detector.

    Source: §I · p. 2, §VIII · p. 11

  • Results

    TEEP-RCNN achieves 73.3% mAP@50 in 10 epochs and scores 0.287 mAP on rolled-in scale versus YOLOv11m’s 0.277 mAP.

    Source: §VIII · p. 11

  • Takeaways & Limitations

    The rolled-in-scale comparison provides evidence of an advantage over YOLOv11m for that class, despite its 100-epoch training run.

    Source: §VIII · p. 11

    Crazing remains difficult: its distributed texture does not fit standard box annotations well, and Faster R-CNN’s anchor-based proposals are poorly suited to it.

    Source: §VII · p. 11

Abstract

from arXiv · show

Steel surface defect detection is critical for automated industrial quality control but remains challenging due to subtle inter-class texture differences and pronounced class imbalance. We introduce TEEP-RCNN (Texture-Enhanced Edge-aware Perception Region-based CNN), a two-stage detector built on Faster R-CNN with a Feature Pyramid Network backbone and an improved Convolutional Block Attention Module (CBAM). Our CBAM adds dropout regularization in the channel attention MLP and batch normalization on the spatial attention branch, reducing co-adaptation and stabilizing gating logits. Training uses a differential learning rate protocol with cosine annealing warm-up, separating update rates for the pre-trained ResNet-101 backbone and the detection head. At inference, predictions are refined via Test-Time Augmentation fused with Weighted Box Fusion (WBF), improving localization stability on elongated and boundary-adjacent defects. On the NEU-DET benchmark across six defect categories, TEEP-RCNN achieves 73.3\% mAP@50 and 37.9\% mAP@50-95 in only 10 training epochs on a single GPU, competitive with YOLOv11m (76.2\% mAP@50, 100 epochs) while outperforming it on the rolled-in-scale category under the COCO metric. Per-class analysis shows the spatial attention branch is most effective on elongated texture defects such as patches and scratches, while crazing remains an open challenge across both paradigms due to its distributed non-local texture structure.

I. INTRODUCTION

Steel defect inspection faces high-stakes quality costs and challenges from texture differences; the paper motivates learned attention to adapt features to distinct defect patterns. It presents TEEP-RCNN with CBAM modifications and training and inference strategies.

I. INTRODUCTION

  • Surface defects contribute to downstream rejects and rework costs; better inspection can prevent material waste and recalls of formed components.The paper describes hot-rolled steel production as exceeding one billion metric tons annually.Source: §I · p. 1
  • Human operators inspect strips moving at 5–30 meters per second, where fatigue-related missed detections are documented.The introduction frames automated vision as a practical necessity.Source: §I · p. 1

A. Traditional Defect Detection

  • Traditional LBP, Gabor, and wavelet descriptors capture surface structure under controlled conditions but require category- and mill-specific tuning and can fail when conditions shift.The paper motivates learning representations directly from annotated examples as a response.Source: §I · p. 1

I. INTRODUCTION

  • Crazing activates across many locations, whereas scratches produce narrow, spatially concentrated responses, motivating feature handling that reflects distinct defect textures.The introduction identifies this mismatch as a gap in standard two-stage detectors.Source: §I · pp. 1–2
  • TEEP-RCNN combines two CBAM modifications with differential-learning-rate training and WBF-TTA inference to address texture-specific detection gaps.The channel MLP adds dropout, and the spatial branch adds batch normalization.Source: §I · p. 2

B. Two-Stage CNN Detectors

This section surveys detector families and attention methods relevant to steel-defect detection, then connects prior industrial findings to the paper’s focus on texture-specific challenges.

B. Two-Stage CNN Detectors

  • Faster R-CNN shares backbone features between proposals and detection, while FPN adds a multi-resolution hierarchy for fine spatial detail and deep semantics.Mask R-CNN extends the family with a pixel-level segmentation branch and RoI Align.Source: B. Two-Stage CNN Detectors · p. 2

C. Single-Stage Detectors

  • SSD, RetinaNet, and YOLO use single-stage detection; RetinaNet addresses foreground-background imbalance, and YOLOv11m is reported at 76.2% mAP@50 on NEU-DET.The excerpt also describes YOLOv8’s anchor-free head and YOLOv11’s C3k2 blocks and SPPF pooling.Source: C. Single-Stage Detectors · pp. 2–3

D. Attention Mechanisms in Detection

  • CBAM sequentially recalibrates channels and spatial activations, and this paper extends it with dropout in the channel MLP and normalization in the spatial branch.The paper identifies these as regularization weaknesses and says the changes retain CBAM’s lightweight design philosophy.Source: D. Attention Mechanisms in Detection · p. 3

E. Transformer-Based and Anchor-Free Detectors

  • Transformer detectors include DETR and Deformable DETR; the cited DSAT result reaches 83.1% mAP@50 on NEU-DET, while TEEP-RCNN uses a 10-epoch CNN training budget.The passage characterizes transformer approaches as requiring larger training budgets and typically larger datasets.Source: E. Transformer-Based and Anchor-Free… · p. 3

F. Deep Learning for Surface Defect Detection

  • Prior NEU-DET work explores deformable convolutions, graph reasoning, and texture-focused augmentation; the paper notes crazing and rolled-in scale are consistently harder than patches and scratches.The paper attributes this pattern to the non-local distributed nature of the former defect types.Source: F. Deep Learning for Surface Defect… · p. 3

III. DATASET AND PREPROCESSING

The study uses NEU-DET’s six defect categories and augmentation tailored to image and surface variation, then applies the data in a Faster R-CNN architecture with multi-scale features.

A. NEU-DET Dataset

  • NEU-DET contains six defect classes, with 1,224 training, 216 validation, and 360 test images after the authors’ split; inclusion has 856 training instances versus 351 for pitted surface.The image size is 200×200 pixels in the database and is resized to 640×640 for the model.Source: A. NEU-DET Dataset · p. 3

B. Data Augmentation

  • Training augmentation spans geometric, photometric, and structural transforms; CLAHE is singled out for enhancing low-contrast pitted-surface cavity boundaries.Coarse and grid dropout simulate partial occlusion, and boxes with less than 30% visibility after transforms are discarded.Source: B. Data Augmentation · p. 4

A. Architecture Overview

  • TEEP-RCNN uses Faster R-CNN with an improved CBAM, morphology-tuned anchors, and WBF-TTA; FPN levels feed the RPN and RoI head, with Soft-NMS and WBF refining detections at inference.The four pyramid levels span strides 4, 8, 16, and 32 relative to the 640×640 input.Source: A. Architecture Overview · p. 4

B. ResNet-101 Backbone with FPN

The detector uses a trainable, ImageNet-pretrained ResNet-101 with an FPN, then applies modified CBAM attention within Layer4 to recalibrate features.

B. ResNet-101 Backbone with FPN

  • The ImageNet-pretrained ResNet-101 has all five residual stage groups trainable and 23 bottleneck blocks in Layer4 for CBAM modulation.The paper contrasts Layer4’s 23 blocks with ResNet-50’s six.Source: B. ResNet-101 Backbone with FPN · p. 4
  • FPN constructs P2–P5 through top-down upsampling and lateral additions, giving 256 channels per level and strides from 4 to 32.P2 preserves finer spatial resolution, while P5 carries coarser, more semantically meaningful features.Source: B. ResNet-101 Backbone with FPN · p. 4, Figure 2 · p. 5

C. Improved CBAM Module

  • The improved CBAM adds dropout at rate 0.1 to the channel-attention MLP and applies channel recalibration before spatial attention.The dropout is intended to prevent consistently routing information through the same dominant channels during training.Source: C. Improved CBAM Module · pp. 4–5
  • Batch normalization before the spatial-attention sigmoid stabilizes logits and keeps gating weights in a soft, gradient-productive regime during training.The passage notes this is especially beneficial on NEU-DET, where batches of eight may contain mostly background images.Source: C. Improved CBAM Module · p. 5

D. Region Proposal Network and Anchor Design

The RPN uses a multi-scale, multi-ratio anchor design, including an added elongated-box template for scratches and rolled-in scale. IoU thresholds label anchors, while proposal filtering limits noise entering the RoI head.

D. Region Proposal Network and Anchor Design

E. RoI Head and Post-Processing

The RoI head uses multi-scale features and class-specific prediction branches, while post-processing combines Soft-NMS and weighted box fusion across augmented views. Training uses separate backbone and head learning rates with warm-up and mixed precision.

E. RoI Head and Post-Processing

  • Multi-Scale RoI Align maps each proposal to a 7 × 7 grid, followed by parallel classification and box-regression branches.The branches produce 7 classification outputs and 24 regression outputs; detections use a 0.5 confidence threshold and per-class NMS threshold of 0.3.Source: E. RoI Head and Post-Processing · pp. 5–6
  • The improved CBAM adds dropout after ReLU in the channel-attention MLP and batch normalization after the spatial-attention convolution.Source: Figure 3 · p. 6
  • Soft-NMS smoothly decays scores for overlapping detections, preserving nearby detections in dense defect regions while allowing a slight increase in false positives.The final score threshold controls the false-positive increase.Source: E. RoI Head and Post-Processing · p. 6
  • WBF merges predictions from the original image and a flipped, slightly scale-jittered view, clustering boxes at IoU ≥0.6 and averaging coordinates by confidence.For elongated defects such as boundary-spanning scratches, coordinate averaging reduces localization variance compared with a single view or hard NMS across views.Source: E. RoI Head and Post-Processing · p. 6

F. Training Strategy

  • The ResNet-101 backbone receives learning rate 10^-4 and the task-specific components 10^-3, a 10:1 ratio intended to protect pretrained representations during early updates.Source: F. Training Strategy · p. 6
  • AdamW uses cosine annealing with a two-epoch linear warm-up; automatic mixed precision enables batch size 8 on a single Tesla T4 GPU.The warm-up limits large early updates before gradient statistics become meaningful.Source: F. Training Strategy · p. 6

V. EXPERIMENTS

Experiments use a shared NEU-DET split and compare Faster R-CNN and YOLO baselines with TEEP-RCNN variants. Models are evaluated using the COCO detection metric suite.

A. Experimental Setup and Baselines

B. Model Efficiency Analysis

TEEP-RCNN has higher reported parameter and compute counts than YOLOv11m, but reaches 70% mAP@50 in fewer training epochs. The passage frames this as a depth-versus-efficiency trade-off relevant to limited data or GPU time.

B. Model Efficiency Analysis

  • 61.8M parameters and approximately 170 GFLOPs at 640×640 are reported for TEEP-RCNN, versus 67.7 GFLOPs for YOLOv11m.The passage attributes TEEP-RCNN’s higher parameter count to its ResNet-101 backbone and FPN head, and its higher FLOPs to backbone depth.Source: B. Model Efficiency Analysis · p. 7
  • TEEP-RCNN reaches 70% mAP@50 in 6 epochs, compared with approximately 50–60 epochs for YOLOv11m.The passage notes this training efficiency matters when labeled data is expensive or GPU time is limited.Source: B. Model Efficiency Analysis · p. 7

C. Quantitative Comparison

TEEP-RCNN delivers strong NEU-DET results with a 10-epoch training budget, while an extended class-weighted run scores lower. Training metrics indicate rapid improvement followed by a late plateau, and results vary by object size.

C. Quantitative Comparison

  • The 200-epoch class-weighted run reaches mAP@50 0.685, below the primary 10-epoch model’s 0.733.The extended run’s best checkpoint was at epoch 42.Source: C. Quantitative Comparison · pp. 7–8
  • TEEP-RCNN reaches mAP@50 0.733 and mAP@50-95 0.379 in 10 epochs, versus YOLOv11m’s 0.762 and 0.450 after 100 epochs.The training budget is 10× smaller than YOLOv11m’s.Source: C. Quantitative Comparison · p. 7, Table V · p. 8
  • At 640-pixel input, mAP medium is 0.307 and mAP large is 0.383; all NEU-DET boxes fall in the medium-to-large range.The source attributes the slightly higher large-object score to stronger activation signals and better overlap with anchor templates.Source: C. Quantitative Comparison · p. 8

D. Training Dynamics

  • Training loss drops from 2.755 in epoch 0 to 0.386 in epoch 1; mAP@50 reaches 0.429 by epoch 3 and exceeds 0.700 by epoch 6.The passage describes loss decreasing alongside rising mAP@50, without loss-mAP decoupling.Source: D. Training Dynamics · p. 8
  • mAP@50 moves from 0.736 at epoch 8 to 0.733 at epoch 9, a plateau the authors interpret as near convergence under this schedule.They report that loss continues to decline and cosine annealing leaves only small, stable weight updates.Source: D. Training Dynamics · p. 8

E. Per-Class Analysis

TEEP-RCNN scores most strongly on patches and scratches, and marginally surpasses YOLOv11m on rolled-in scale despite a shorter training run. Crazing is the weakest class for both models, with distributed texture posing a localization challenge.

E. Per-Class Analysis

  • TEEP-RCNN’s strongest per-class scores are patches (0.537) and scratches (0.486).The passage associates patch activations with spatially contiguous regions and scratch activations with narrow, directional lines.Source: E. Per-Class Analysis · p. 8, Table VII · p. 9
  • Inclusion (0.406) and pitted surface (0.457) have mid-range scores, consistent with their compact but visually distinctive morphology.Source: E. Per-Class Analysis · p. 8, Table VII · p. 9
  • On rolled-in scale, TEEP-RCNN (0.287) marginally outperforms YOLOv11m (0.277), despite a 10:1 training-epoch ratio.The passage relates the result to irregular elongated regions and the spatial attention branch’s 7×7 receptive field.Source: E. Per-Class Analysis · pp. 8–9, Figure 5 · p. 9
  • Crazing has the lowest per-class scores for both TEEP-RCNN (0.099) and YOLOv11m (0.148), with distributed crack texture challenging localized boxes.The passages describe full-image activations and a mismatch between distributed texture and tight bounding-box annotations.Source: E. Per-Class Analysis · pp. 8–9, Figure 5 · p. 9

F. Qualitative Detection Results

Qualitative examples show TEEP-RCNN localizing several defect types, while crazing remains prone to overlapping proposals. Discussion connects fast convergence and class-specific behavior to the model’s attention and training choices, and identifies localization precision as a boundary.

F. Qualitative Detection Results

  • Crazing images show over-proposal across crack networks, aligning with low per-class mAP and indicating a localization-precision challenge rather than class discrimination.Source: F. Qualitative Detection Results · p. 9

VI. DISCUSSION

  • The model reaches mAP@50 0.700 by epoch 6 and 0.733 by epoch 9, using 10× fewer training epochs than YOLOv11m.The discussion attributes convergence efficiency to ImageNet-pretrained ResNet-101 features and targeted attention.Source: §VI · p. 9
  • Per-class results indirectly suggest 7×7 spatial attention responds to elongated textures, while crazing’s diffuse activations provide little localization signal.Explicit attention-map visualization is left for future work.Source: §VI · p. 9
  • The 200-epoch class-weighted run reaches mAP@50 0.685 at epoch 42, below the primary model’s 0.733; the discussion cautions against weighting by class frequency alone.It proposes weighting in proportion to each class’s trainable capacity.Source: §VI · pp. 9–10

F. Qualitative Detection Results

  • Examples show localization of scratches, crazing regions, and multiple rolled-in-scale instances; prediction confidence scores range from 0.44 to 0.99.The figure contrasts solid green ground-truth boxes with dashed red predictions.Source: F. Qualitative Detection Results · p. 9, Figure 6 · p. 10

VI. DISCUSSION

  • TEEP-RCNN’s mAP@50-95 of 0.379 trails YOLOv11m’s 0.450; the authors partly attribute the gap to the 10-epoch duration limiting regression-head convergence.They note that tight box coordinates require learning fine-grained offsets from anchor templates.Source: §VI · p. 10

F. Qualitative Detection Results

  • Across the extended run, patches and scratches stabilize early, rolled-in scale remains variable, and crazing shows no resolved localization challenge.The heatmap uses warm colors for higher mAP; crazing remains consistently blue.Source: F. Qualitative Detection Results · p. 9, Figure 7 · p. 10

VI. DISCUSSION

  • For offline quality control, the discussion prioritizes detection accuracy and false-negative rate over inference speed and values an auditable proposal-then-classify pipeline.It notes that the RPN score distribution can be monitored separately from classification confidence.Source: §VI · pp. 10–11

VII. LIMITATIONS

On NEU-DET, TEEP-RCNN approaches YOLOv11m’s mAP@50 with far fewer training epochs and outperforms it on rolled-in-scale. The conclusion highlights its attention and inference strategy alongside limitations in model size, class coverage, ablation evidence, dataset scope, and latency measurement.

VII. LIMITATIONS

  • At 61.8 million parameters versus YOLOv11m’s 20.1 million, TEEP-RCNN would need a lighter backbone for edge deployment, where improved CBAM’s impact remains open.ResNet-101 depth accounts for most of the parameter overhead.Source: §VII · p. 11
  • Crazing reaches only 0.099 per-class mAP; bounding boxes and Faster R-CNN proposals are poorly suited to its distributed texture without a compact centroid.The conclusion suggests a segmentation auxiliary branch or texture-specific detection head for this class.Source: §VII · p. 11, §VIII · p. 11
  • Without a controlled standard-versus-improved CBAM ablation, the separate effects of dropout and batch normalization remain unisolated.The authors leave an ablation requiring at least four experimental runs to future work.Source: §VII · p. 11
  • All experiments use NEU-DET, so generalizability to other industrial defect datasets has not been verified.The authors note behavior may differ with finer resolution, more complex backgrounds, or more defect classes.Source: §VII · p. 11
  • Inference throughput was not measured, leaving latency comparison on Tesla T4 unestablished for deployment parity with YOLO-based solutions.The authors state a rigorous platform-specific latency comparison would be necessary before making that claim.Source: §VII · p. 11

VIII. CONCLUSION

  • TEEP-RCNN combines improved CBAM, differential learning rates with cosine annealing warm-up, and WBF-TTA for more consistent localization on elongated defects.The CBAM adds channel-MLP dropout and spatial-branch batch normalization.Source: §VIII · p. 11
  • 73.3% mAP@50 in 10 epochs places TEEP-RCNN within 3 percentage points of YOLOv11m trained for 100 epochs; rolled-in-scale mAP is 0.287 versus 0.277.The authors connect the rolled-in-scale result to spatial attention’s suitability for elongated defect morphology.Source: §VIII · p. 11

Related papers

Loading 2609.28077v1…