An Improved Convolutional Neural Network System for Automatically Detecting Rebar in GPR Data

Abbas Rashidi; Ge (Gaby) Ou; Zhongming Xiang

REVIEW 3 major objections 2 minor 12 references

An Improved Convolutional Neural Network System for Automatically Detecting Rebar in GPR Data

T0 review · 3 major / 2 minor · reviewed 2026-05-24 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AlexNet detects rebar in GPR scans more accurately than a basic CNN, especially when the bars sit close together.

desk verdict AlexNet beats a baseline CNN on GPR rebar detection but the comparison lacks matched training controls so architecture may not explain the gains. read the letter →

arxiv 1907.09997 v1 pith:MA3I6FDB submitted 2019-07-23 cs.CV

Zhongming Xiang , Abbas Rashidi , Ge (Gaby) Ou This is my paper

classification cs.CV

keywords rebardetectiongroundpenetratingradarGPRAlexNetconvolutionalneuralnetworkconcretestructuresimageclassificationnon-destructivetesting

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper applies AlexNet, a convolutional neural network already successful at general image tasks, to the problem of locating steel reinforcing bars inside concrete using ground-penetrating radar images. It builds a side-by-side comparison with a simpler CNN and tests both on scans that differ in how densely the bars are spaced and how large each analysis window is cut. The central finding is that AlexNet delivers higher accuracy overall, and the gap widens precisely when bars are packed tightly and signals begin to overlap. Accuracy also rises or falls sharply with window size, because the window must contain enough of the characteristic rebar reflection pattern. Uniform, widely spaced bar layouts turn out easier to classify than crowded or irregular ones because they produce less signal interference.

What carries the argument

AlexNet applied to fixed-size windows cut from GPR B-scans, used to classify whether each window contains a rebar signature, compared directly against a baseline convolutional network on the same task.

What would settle it

Train both networks from scratch on exactly the same GPR image set with identical preprocessing and hyperparameter search, then measure whether the accuracy gap on dense-mesh test cases disappears.

Watch

Extended reading notes

Core claim

AlexNet outperforms a traditional CNN for rebar detection in GPR data, with the performance margin largest on densely distributed meshes. Detection accuracy changes markedly with the size of the image window used for classification, and windows must be large enough to capture adequate rebar signal information. Uniformly and sparsely spaced rebar meshes produce clearer results than dense or uneven arrangements because they reduce the chance of overlapping radar reflections.

Load-bearing premise

Any accuracy difference between AlexNet and the traditional CNN arises from the choice of network architecture rather than unequal training data sizes, preprocessing steps, or hyperparameter choices.

Editorial extensions

If this is right

Automated rebar mapping becomes more reliable in crowded reinforcement layouts where manual interpretation is hardest.
Window size must be chosen to enclose a complete rebar reflection pattern or accuracy drops.
Sparse and regular bar grids remain easier targets even for the improved network.
Signal overlap from nearby bars is the main factor limiting performance on dense meshes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same window-based classification approach could be tested on other embedded objects such as conduits if their radar signatures are similarly localized.
Once trained, the model could be run on streaming GPR data collected during routine bridge or building scans to flag rebar locations in real time.
Collecting more examples of uneven or clustered rebar layouts might close the remaining accuracy gap on those harder cases.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The main point is that this paper takes AlexNet and applies it to spotting rebar in GPR images, reporting better accuracy than a basic CNN especially when the rebar is packed close together. It also checks how rebar spacing and the size of the analysis window affect results. That part is straightforward and useful for the domain. They correctly note that dense or uneven meshes cause more signal overlap and are harder to read, and that the window needs to be large enough to capture the rebar signature without too much extra noise. Those observations line up with how GPR data actually behaves. The comparison itself is the soft spot. The abstract gives no sign that both models saw the same data volume, the same preprocessing steps, or the same level of hyperparameter work. If the traditional CNN was trained with less effort or smaller effective sets, the reported edge for AlexNet cannot be pinned on the architecture. Dataset size, splits, and any error bars or statistical checks are also absent, which makes the numbers hard to trust or repeat. This is aimed at civil engineers who already work with GPR for infrastructure checks. Someone in that niche could use the window-size and density findings as a starting point. A reader from computer vision or machine learning will see only a routine transfer of a 2012 model to a new image set. The paper shows honest engagement with the application problem but the empirical claims need tighter controls before they can be taken at face value. It is worth sending to peer review so the methods section can be examined directly.

Referee Report

3 major / 2 minor

Summary. The paper proposes AlexNet as an improved CNN for automatic rebar detection in GPR images of concrete structures. It constructs a traditional CNN as baseline, evaluates effects of rebar mesh density (dense/sparse/uniform/uneven) and splitting window size on accuracy, and reports that AlexNet outperforms the baseline (with larger gains on dense meshes), that accuracy varies with window size, and that uniform/sparse arrangements are easier to detect due to reduced signal interference.

Significance. If the reported accuracy differences can be shown to arise from architecture rather than unequal training regimes, the work would provide a practical demonstration of deep CNNs for GPR-based rebar inspection, with potential value for civil-engineering quality control where dense rebar is common. The empirical focus on real-world arrangement variations is a modest strength, but absence of dataset scale, splits, or error bars keeps the result from being immediately usable or citable as a benchmark.

major comments (3)

[Experimental results / Section 3] Experimental results (and Section 3): the manuscript supplies no information on dataset size, number of GPR images, train/validation/test splits, or whether the same data volume and preprocessing pipeline were used for both AlexNet and the traditional CNN. Without these controls the headline claim that AlexNet outperforms the baseline (especially on dense meshes) cannot be verified and may be confounded by training-regime differences.
[Experimental results] Experimental results: no hyperparameter search protocol, early-stopping criteria, or optimization effort is reported for the traditional CNN baseline. If the baseline received less tuning than AlexNet, the observed superiority cannot be attributed to architecture.
[Abstract / results] Abstract and results: no error bars, statistical significance tests, or cross-validation details accompany the accuracy comparisons, so the statement that superiority is 'more notable' for dense meshes lacks quantitative support.

minor comments (2)

Notation for window size and rebar density categories is introduced without a clear table or figure legend, making it hard to reproduce the exact experimental conditions.
The abstract states comparative results but the full text should explicitly list the traditional CNN architecture (layer counts, filter sizes) for direct comparison with AlexNet.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on the experimental reporting in our manuscript. We address each major point below and will revise the manuscript to incorporate the requested details.

read point-by-point responses

Referee: [Experimental results / Section 3] Experimental results (and Section 3): the manuscript supplies no information on dataset size, number of GPR images, train/validation/test splits, or whether the same data volume and preprocessing pipeline were used for both AlexNet and the traditional CNN. Without these controls the headline claim that AlexNet outperforms the baseline (especially on dense meshes) cannot be verified and may be confounded by training-regime differences.

Authors: We agree that these details are essential for verifying the comparison. The same GPR dataset and preprocessing pipeline were applied to both models. We will add the dataset size, number of images, and explicit train/validation/test splits to the revised Section 3 to confirm the data volumes were identical. revision: yes
Referee: [Experimental results] Experimental results: no hyperparameter search protocol, early-stopping criteria, or optimization effort is reported for the traditional CNN baseline. If the baseline received less tuning than AlexNet, the observed superiority cannot be attributed to architecture.

Authors: This criticism is valid. We will expand the experimental section to describe the hyperparameter search protocol, early-stopping criteria, and optimization settings used for the traditional CNN, showing that comparable effort was applied to both architectures. revision: yes
Referee: [Abstract / results] Abstract and results: no error bars, statistical significance tests, or cross-validation details accompany the accuracy comparisons, so the statement that superiority is 'more notable' for dense meshes lacks quantitative support.

Authors: We accept that error bars and statistical tests are needed to support the claims. In the revision we will add cross-validation results, error bars, and significance tests to the accuracy comparisons, particularly to quantify the differences observed on dense meshes. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely empirical model comparison with no derivation chain

full rationale

The paper is an empirical study comparing AlexNet against a baseline CNN on GPR rebar detection tasks. No equations, first-principles derivations, predictions, or ansatzes appear in the abstract or described content. Claims rest on experimental accuracy metrics rather than any reduction of outputs to inputs by construction. No self-citation load-bearing steps or uniqueness theorems are invoked. This matches the default case of a self-contained empirical paper; the noted weakness in training controls is a validity concern, not circularity.

Assumptions & free parameters 2 free parameters · 1 assumptions · 0 invented entities

The work rests on the pre-existing AlexNet architecture and standard supervised image-classification assumptions; no new entities are introduced and the only free parameters are implicit training choices not detailed in the abstract.

free parameters (2)

window size
Multiple window sizes are tested but the selection process and any optimization are not described.
model hyperparameters
Training settings for both AlexNet and the traditional CNN are unspecified in the provided text.

assumptions (1)

domain assumption GPR B-scan images can be treated as ordinary 2D photographs for direct input to an ImageNet-pretrained CNN without domain-specific feature engineering.
Implicit in the decision to apply AlexNet directly to the GPR data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Improved Convolutional Neural Network System for Automatically Detecting Rebar in GPR Data." pith.science (2026). https://pith.science/paper/MA3I6FDB

@misc{pith2026190709997,
  author       = {Pith},
  title        = {Pith review of: An Improved Convolutional Neural Network System for Automatically Detecting Rebar in GPR Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MA3I6FDB}},
  note         = {Machine review of arXiv:1907.09997}
}

read the original abstract

As a mature technology, Ground Penetration Radar (GPR) is now widely employed in detecting rebar and other embedded elements in concrete structures. Manually recognizing rebar from GPR data is a time-consuming and error-prone procedure. Although there are several approaches to automatically detect rebar, it is still challenging to find a high resolution and efficient method for different rebar arrangements, especially for closely spaced rebar meshes. As an improved Convolution Neural Network (CNN), AlexNet shows superiority over traditional methods in image recognition domain. Thus, this paper introduces AlexNet as an alternative solution for automatically detecting rebar within GPR data. In order to show the efficiency of the proposed approach, a traditional CNN is built as the comparative option. Moreover, this research evaluates the impacts of different rebar arrangements and different window sizes on the accuracy of results. The results revealed that: (1) AlexNet outperforms the traditional CNN approach, and its superiority is more notable when the rebar meshes are densely distributed; (2) the detection accuracy significantly varies with changing the size of splitting window, and a proper window should contain enough information about rebar; (3) uniformly and sparsely distributed rebar meshes are more recognizable than densely or unevenly distributed items, due to lower chances of signal interferences.

Figures

Figures reproduced from arXiv: 1907.09997 by the authors.

**Figure 1.** The network structure of AlexNet As a deep learning method, AlexNet incorporates several new features into CNN. These additions improve both recognition accuracy and computation efficiency (Krizhevsky et al. 2012). Different features of AlexNet, compared to traditional CNN, are summarized in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** , there are four typical shapes in the segmented images: left, peak, right, and other. If the shape in the sub-image is notably left, peak or right, it will be correspondingly classified as ‘left’, ‘peak’ or ‘right’. Otherwise, and if the segment image does not include any recognizable shape, it will be classified as ‘other’. Meanwhile, based on the requirement of input data, all of these images are resized to 227×2… view at source ↗

**Figure 3.** Examples of different window sizes: a-120×30; b-150×50; b-200×80; d-250×100 EXPERIMENTAL SETUP AND CASE STUDY For demonstrating the efficiency of the proposed AlexNet system, a number of experiments have been conducted. This section briefly reviews the selected experimental setup as well as the obtained results. Experimental Data Several reinforced concrete elements in a newly renovated building have been selected a… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Three Case studies used as testbed: a-Column; b-Shear Wall; c-Suspended Slab [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: A typical GPR imagery data obtained by scanning a sample concrete element Results and Discussion The purpose of using the trained AlexNet and TraNet systems is to predict type of label in each small part [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

E., & Stimac, P

Besaw, L. E., & Stimac, P. J. (2015, May). Deep convolutional neural networks for classifying GPR B- scans. In Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XX (Vol. 9454, p. 945413). International Society for Optics and Photonics

work page 2015
[2]

O., & Roska, T

Chua, L. O., & Roska, T. (1993). The CNN paradigm. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 40(3), 147-156

work page 1993
[3]

Dinh, K., Gucunski, N., & Duong, T. H. (2018). An algorithm for automatic localization and detection of rebars from GPR data of concrete bridge decks. Automation in Construction, 89, 292-298

work page 2018
[4]

R., & Cohn, A

Dou, Q., Wei, L., Magee, D. R., & Cohn, A. G. (2017). Real-time hyperbola recognition and fitting in GPR data. IEEE Transactions on Geoscience and Remote Sensing, 55(1), 51-62

work page 2017
[5]

J., Chiou, C

Eisenmann, D., Margetan, F. J., Chiou, C. P., Ellis, S., Huang, T., & Tan, J. Y. (2017, February). Effects of position, orientation, and metal loss on GPR signals from structural rebar. In AIP Conference Proceedings (Vol. 1806, No. 1, p. 080005). AIP Publishing

work page 2017
[6]

Guyon, I., & Elisseeff, A. (2006). An introduction to feature extraction. In Feature extraction (pp. 1-25). Springer, Berlin, Heidelberg

work page 2006
[7]

J., Romero, F

Kaur, P., Dana, K. J., Romero, F. A., & Gucunski, N. (2016). Automated GPR rebar analysis for robotic bridge deck evaluation. IEEE transactions on cybernetics, 46(10), 2265-2276

work page 2016
[8]

Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105)

work page 2012

Show all 12 references

[9]

(2017, August)

Lameri, S., Lombardi, F., Bestagini, P., Lualdi, M., & Tubaro, S. (2017, August). Landmine detection from GPR data using convolut ional neural networks. In Signal Processing Conference (EUSIPCO), 2017 25th European (pp. 508-512). IEEE

2017
[10]

L., & Mokji, M

Lee, K. L., & Mokji, M. M. (2014, August). Automatic target detection in GPR images using Histogram of Oriented Gradients (HOG). In Electronic Design (ICED), 2014 2nd International Conference on (pp. 181-186). IEEE

2014
[11]

& Berg, A

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Berg, A. C. (2015). Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252

2015
[12]

Yuan, C., Li, S., Cai, H., & Kamat, V. R. (2018). GPR Signature Detection and Decomposition for Mapping Buried Utilities with Complex Spatial Configuration. Journal of Computing in Civil Engineering, 32(4), 04018026

2018

Pith tools

Reviewed May 24, 2026 · model on record in the stance chip above.

[1] [1]

E., & Stimac, P

Besaw, L. E., & Stimac, P. J. (2015, May). Deep convolutional neural networks for classifying GPR B- scans. In Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XX (Vol. 9454, p. 945413). International Society for Optics and Photonics

work page 2015

[2] [2]

O., & Roska, T

Chua, L. O., & Roska, T. (1993). The CNN paradigm. IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 40(3), 147-156

work page 1993

[3] [3]

Dinh, K., Gucunski, N., & Duong, T. H. (2018). An algorithm for automatic localization and detection of rebars from GPR data of concrete bridge decks. Automation in Construction, 89, 292-298

work page 2018

[4] [4]

R., & Cohn, A

Dou, Q., Wei, L., Magee, D. R., & Cohn, A. G. (2017). Real-time hyperbola recognition and fitting in GPR data. IEEE Transactions on Geoscience and Remote Sensing, 55(1), 51-62

work page 2017

[5] [5]

J., Chiou, C

Eisenmann, D., Margetan, F. J., Chiou, C. P., Ellis, S., Huang, T., & Tan, J. Y. (2017, February). Effects of position, orientation, and metal loss on GPR signals from structural rebar. In AIP Conference Proceedings (Vol. 1806, No. 1, p. 080005). AIP Publishing

work page 2017

[6] [6]

Guyon, I., & Elisseeff, A. (2006). An introduction to feature extraction. In Feature extraction (pp. 1-25). Springer, Berlin, Heidelberg

work page 2006

[7] [7]

J., Romero, F

Kaur, P., Dana, K. J., Romero, F. A., & Gucunski, N. (2016). Automated GPR rebar analysis for robotic bridge deck evaluation. IEEE transactions on cybernetics, 46(10), 2265-2276

work page 2016

[8] [8]

Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105)

work page 2012