TileNet: Tile-Based CNN-SVM Architecture for Autonomous Unmanned Aerial Systems Inspection of Flat Roofs
Samuel Dunthorne, Hashim A. Hashim
cs.CV · cs.AI
Quick verdict
Claims 94.4% roof-defect accuracy from a lightweight CNN-SVM on UAS tiles; open data strong, but roof-level leakage and dated baselines temper the result.
Summary
The paper presents TileNet, a tile-based CNN-SVM pipeline for binary (defect / no-defect) classification of modified-bitumen flat-roof tiles from UAS imagery. High-resolution photographs (cropped to 3840×2160) are split into 640×360 tiles, labeled by a single annotator with five years at a roofing company, augmented offline, and classified by a five-convolution, four-dense network whose final layer is a linear squared-hinge SVM head trained with SGD (momentum 0.2). On a photo-level split (43,383/3,869/2,540 train/validation/test tiles) it reaches 94.4% mean test accuracy (95% CI ±0.4% over three seeds), recall 0.920, precision 0.956, average precision 0.980, and classifies a 36-tile image in 7.8 ms on an RTX 5070 — ahead of GoogLeNet (89.2% ±5.7%) and AlexNet (79.8% ±14.0%). Dataset, scripts, and weights are openly archived on Zenodo. Main limitations, mostly acknowledged by the authors: overlapping photographs of the same roof cross splits, so roof-level leakage likely inflates accuracy; baselines are dated with wide confidence intervals and no modern lightweight architecture is compared; inference was verified only on a workstation GPU, not onboard; the dual-altitude strategy is not ablated; Adam collapses in a single untuned run; and labels lack inter-annotator agreement.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A tile-based CNN-SVM (five convolutional, four dense layers, linear squared-hinge head) achieves 94.4% ±0.4% mean test accuracy for tile-level binary defect classification of modified-bitumen roof imagery on a photo-level split.
- The model outperforms GoogLeNet (89.2% ±5.7%) and AlexNet (79.8% ±14.0%) on identical data and splits while classifying a 36-tile image in 7.8 ms on an RTX 5070, versus 38.9 ms and 22.5 ms for the baselines.
- SGD with momentum is essential to training the architecture — Adam reportedly collapses to the majority class — and offline data augmentation curbs overfitting.
- The framework supports a real-time dual-altitude UAS inspection workflow, though only the per-tile classification stage is implemented and benchmarked.
Strengths
Open artifacts with a fixed split manifest
Dataset with manifest.csv, tiling/augmentation/training/evaluation scripts, and final weights are CC BY 4.0 archived on Zenodo (§V-A); seeds are fixed, enabling exact re-runs — unusual completeness for this application domain.
Explicit honest scoping of what is implemented
§IV-B states plainly that only the tile-level classifier, training pipeline, metrics, and inference timing are implemented and benchmarked; closed-loop flight, onboard execution, georeferenced mapping, and revisit waypoints are future work. §V-E pre-registers the leakage concern.
Systematic architecture sweep plus targeted ablations
Sound split and metric hygiene
Internally consistent quantitative bookkeeping
Weaknesses
Roof-level leakage plausibly inflates the headline number
Dated baselines misaligned with the efficiency thesis
Single-run optimizer conclusion
Deployment claims rest on workstation measurements
Label and provenance gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
TileNet (arXiv:2609.13013) — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does TileNet achieve for roof defect detection?
On its held-out test set (2,540 tiles from a photo-level split), TileNet reports 94.4% mean test accuracy (95% CI ±0.4% over three seeds), with recall 0.920 and precision 0.956 for the defect class. It is a tile-level binary classifier, not pixel-level segmentation.
Is the TileNet dataset or code publicly available?
Yes. The authors archive the tiled/augmented dataset with a train/val/test split manifest, the processing and training scripts, and the final model weights on Zenodo under CC BY 4.0 (dataset DOI 10.5281/zenodo.20594605; weights 10.5281/zenodo.20596485; note the paper cites two different DOIs for the scripts, 20599364 and 20599363).
Does TileNet run onboard the drone in real time?
Not yet. All reported inference speeds (7.8 ms per 36-tile image, roughly 130 images/s) were measured on an RTX 5070 workstation GPU. Onboard execution and power/thermal profiling on embedded accelerators such as NVIDIA Jetson Orin are explicitly listed as future work, and the FP32 model weighs about 446 MB.
Which roof types does TileNet work on?
Only modified-bitumen (mod-bit) flat-roof membranes, the system type used to build the labeled dataset. The authors explicitly do not claim generalization to TPO, EPDM, or PVC systems, and note that environmental confounders like rain-wetted membranes remain a known source of ambiguity.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The abstract supplied in the task metadata ends mid-word ('Experiment'), but the complete abstract appears at the start of the embedded paper text; the body runs through Appendix A and reference [60] with no visible truncation, so no content-critical loss is assumed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.