Current title:
"Netra AI: A Cascading Multi-Tier Deep Learning Architecture for Automatic Number Plate Recognition in Indian University Campus Surveillance Environments"
Problem: 27 words. LNCS titles are typically 8–14 words. The methodology ("Cascading Multi-Tier"), the application domain ("Indian University Campus"), and the environment ("Surveillance Environments") are all crammed in. Reviewers will see this as unfocused.
Suggested alternatives (pick one):
| Option | Title | Words |
|---|---|---|
| A | Ensemble OCR with Grammar-Aware Post-Processing for Indian License Plate Recognition | 11 |
| B | Multi-Variant Preprocessing and Ensemble OCR for Robust Indian ANPR | 10 |
| C | Robust License Plate Recognition via Confidence-Weighted OCR Ensemble and Positional Grammar Correction | 13 |
| D | Confidence-Weighted Ensemble OCR for Indian License Plate Recognition Under Adverse Conditions | 12 |
Tip
Option A is the strongest. It names the two core novelties (ensemble OCR + grammar post-processing) and the domain (Indian plates) without overpromising. The campus/surveillance/edge details belong in the abstract, not the title.
Current text (page 3):
"both domains require positional grammar constraints to guide character-level decisions, explicit disambiguation of visually confusable characters arising from font and print quality variation, and domain-specific post-processing to enforce structural validity."
Problem: You list only 3 challenges but have 6 contributions (i)–(vi). There is no one-to-one mapping from "gap identified" → "contribution that fills it." Reviewers will ask: "Which of your contributions addresses which gap?"
Suggested rewrite — expand challenges and explicitly number-map them:
License plate recognition in Indian surveillance environments presents five
interconnected challenges that are not simultaneously addressed by any prior
method:
(C1) Imaging degradation diversity: campus gate cameras capture plates under
direct sunlight, nighttime with headlight glare, motion blur, and low-light
dawn conditions, requiring degradation-robust preprocessing rather than
single-condition optimisation → addressed by contribution (i).
(C2) Character ambiguity under Indian plate typography: the standard Indian
plate font yields systematic confusions (O/D/0, I/1/L, B/8, S/5, G/6)
that cannot be resolved by OCR confidence alone and require positional
grammar constraints → addressed by contributions (ii) and (iv).
(C3) Two-line motorcycle plate format: Indian two-wheelers use vertically
stacked dual-line plates that standard single-line OCR pipelines fail to
parse, and no prior Indian ANPR method handles these systematically →
addressed by contribution (iii).
(C4) Structural validation against Indian registration grammar: Indian plates
follow a positional format (2 letters + 1–2 digits + 1–3 letters + 1–4
digits) with 38 valid RTO state codes, but no existing system enforces
this grammar comprehensively → addressed by contribution (iv).
(C5) Edge deployability: campus surveillance requires real-time operation on
resource-constrained hardware (e.g., Raspberry Pi 5), which constrains
model size and inference strategy → addressed by contribution (v).
Important
This creates a clean C1→(i), C2→(ii)+(iv), C3→(iii), C4→(iv), C5→(v) mapping. Contribution (vi) is a result claim, not a gap-filler, so it should stay as a results contribution rather than pretending to fill a gap.
Problem: The dual-strip decomposition is listed as contribution (iii) — a key novelty — but the supporting evidence is:
- Bikes: 93.55% on n=31
- Scooty: 81.82% on n=11
At n=11, the 95% Wilson confidence interval for 81.82% is [52.3%, 94.9%] — a 42-point range. This is not statistically meaningful. A reviewer will immediately flag this.
Two options:
Collect at least 50 additional two-wheeler captures (target: ~100 total). This narrows the CI to a credible range and makes the contribution defensible.
Rewrite contribution (iii) and the relevant Table 5 discussion:
- (iii) An aspect-ratio-aware dual-strip decomposition strategy for two-line
- motorcycle registration plates [...] addressing a vehicle class
- systematically neglected in prior Indian ANPR literature.
+ (iii) An aspect-ratio-aware dual-strip decomposition strategy for two-line
+ motorcycle registration plates [...]. While the two-wheeler subset
+ (n=42) is sufficient to demonstrate the mechanism, the scooter
+ sub-category (n=11) is too small for statistically robust conclusions;
+ we report these results for completeness and note that larger-scale
+ two-wheeler evaluation is needed to confirm the observed accuracy
+ differential.And in Section 5.5 discussion:
- The dual-strip decomposition strategy is responsible for successful
- recognition in 29 out of 31 captures.
+ The dual-strip decomposition strategy is responsible for successful
+ recognition in 29 out of 31 motorcycle captures. The scooter accuracy
+ of 81.82% (9/11) should be interpreted with caution given the small
+ sample size (Wilson 95% CI: [52.3%, 94.9%]); the lower accuracy is
+ attributable to the smaller physical plate size and greater rear
+ bodywork occlusion rather than a decomposition failure.Warning
If you keep the current wording without qualification, this will be flagged by reviewers. Option B is the minimum fix.
Current: † ‡ § ⋆ footnotes scattered through Table 1.
Problem: Four different symbols is too many for a single table. The reader has to constantly look down to decode what each symbol means. In a dense comparison table, this creates cognitive overload.
Suggested fix — inline the qualifications directly into the cells:
- 88.5§
+ 88.5 (91.0 w/ heuristics)
- ∼4.5‡
+ ∼4.5 (CCPD-trained)
- 95.82⋆
+ 95.82 (3 states: GJ, HR, MH)Then remove the footnote legend entirely. If space is tight, use a single footnote * for the one most important qualifier and inline the rest.
Alternatively, add a "Notes" column to Table 1 instead of footnotes:
| Method | Acc. (%) | ... | Notes |
|---|---|---|---|
| Ravirathinam & Patawari [7] | 88.5 | ... | 91.0 with heuristics |
| Nadiminti et al. [10] | ~4.5 | ... | CCPD-trained; fails on Indian |
| Agrawal & Joshi [11] | 95.82 | ... | 3 states only (GJ, HR, MH) |
Current heading:
"Comparison of proposed system with existing Indian ANPR methods. Full-plate Acc.: exact-match accuracy (%). CER: Character Error Rate (%). Ensemble: multiple OCR engines with confidence scoring. Adverse: evaluated under nighttime / motion-blur / low-light. RTO Grammar: 38 Indian state codes with positional correction. †Char-level only; full-plate match not reported. ‡CCPD-trained; fails on Indian plates. §Base 88.5%; 91.0% with heuristics. ⋆3 states only (GJ, HR, MH)."
Problem: This is 85 words. LNCS table captions should be ≤ 30 words. The abbreviation definitions and footnotes are bloating it.
Suggested rewrite:
Table 1: Feature comparison with existing Indian ANPR methods. Accuracy
figures are from respective original evaluations on different datasets; direct
cross-dataset comparison is not possible (see text).
Move all abbreviation definitions and footnotes to the table body or to a paragraph immediately following the table.
Current claim:
"surpassing the prior best by 1.50 percentage points"
Problem: This is a central claim of the paper, but:
- Your 97.32% is on CSD-747 (campus gate, mixed vehicles, 38 RTO codes)
- Agrawal & Joshi's 95.82% is on a different private dataset (commercial trucks, 3 states only)
- Different cameras, locations, vehicle mixes, lighting, plate diversity
The 1.50 pp gap is not a controlled comparison. A reviewer will reject the "outperforming" claim.
Suggested fix — reframe throughout the paper:
In Section 2.4 (Table 1 caption), you already say:
"Direct cross-dataset comparison is not possible as no shared public
Indian ANPR benchmark exists"
BUT then in the abstract, contributions, and conclusion you say:
- "outperforming six existing Indian ANPR methods by at least 1.50
- percentage points"
+ "achieving 97.32% full-plate accuracy on a 747-image campus surveillance
+ dataset, exceeding the highest previously reported Indian full-plate
+ accuracy of 95.82% [11] while noting that direct comparison is
+ cross-dataset and reflects different evaluation conditions"In contribution (vi):
- (vi) Achieve state-of-the-art full-plate accuracy of 97.32% [...],
- outperforming six existing Indian ANPR methods by at least 1.50
- percentage points
+ (vi) Achieve full-plate accuracy of 97.32% on 747 real-world captures
+ spanning 38 RTO codes and diverse conditions, establishing a new
+ reference point for Indian ANPR evaluation. While the highest
+ previously reported accuracy is 95.82% [11], this comparison is
+ cross-dataset and the improvement should be interpreted in that
+ context.Caution
The word "outperforming" implies a controlled comparison. Replace with "exceeding previously reported accuracy" or "establishing a new reference point." This is the single change most likely to prevent a reviewer rejection.
Current formula:
S(v) = 1000 · c_ocr · (0.5 + sim) + |t| if valid
100 · c_ocr · (0.5 + sim) + |t| otherwise
Problem: The 1000× vs 100× multiplier and the 0.5 offset appear without justification. A reviewer will ask:
- Why 10× penalty and not 5× or 100×?
- Why
0.5 + simand not justsim? - Were these tuned on the test set? (data leakage concern)
Suggested fix — add a paragraph after Equation (1):
The 10× penalty ratio (1000 vs. 100) ensures that any format-valid hypothesis
is preferred over any format-invalid hypothesis when OCR confidence is
comparable, encoding the domain prior that structurally valid Indian plates
are overwhelmingly more likely to be correct. The 0.5 offset in the similarity
term (0.5 + sim) prevents the score from collapsing to zero when inter-engine
agreement is low (sim → 0), which occurs frequently for degraded captures
where only one engine produces a usable output; without this offset, valid
single-engine results would be suppressed. These values were selected on a
held-out development set of 50 captures (disjoint from the 747-image evaluation
set) and remained fixed throughout all reported experiments. Sensitivity
analysis (Appendix A) confirms that accuracy varies by less than 0.5 pp for
penalty ratios in [5×, 50×] and offsets in [0.3, 0.7].
Important
If you did NOT tune on a separate dev set, you should either (a) actually run a quick sensitivity sweep now and add it, or (b) honestly state that the values were set heuristically based on the structure of the scoring problem. Do NOT claim dev-set tuning if it didn't happen.
Current text is a single dense paragraph. Rewrite as:
5.1 Evaluation Metrics
The following metrics are used to evaluate system performance:
- **Full-plate exact-match accuracy (Acc.)**: The proportion of test images
where the predicted plate string exactly matches the ground truth, with no
character errors tolerated. This is the primary metric, as partial matches
are operationally useless in surveillance applications.
- **Character Error Rate (CER)**: The Levenshtein edit distance between the
predicted and ground truth strings, normalised by the ground truth string
length. CER quantifies the severity of recognition errors: a CER of 0%
indicates perfect character-level recognition, while values below 1%
indicate predominantly single-character substitutions.
- **Detection Precision (P)**: The fraction of detected plate regions that
correspond to actual license plates. P = TP / (TP + FP).
- **Detection Recall (R)**: The fraction of actual license plates that were
successfully detected. R = TP / (TP + FN). In this evaluation, all plates
present in the evaluation set were detected (FN = 0), yielding R = 1.000.
- **F1-score**: The harmonic mean of Precision and Recall,
F1 = 2PR / (P + R), providing a single balanced metric for detection
performance.
A recognition is counted as a True Positive (TP) if the plate is both
localised by the detector and the recognised string exactly matches the
ground truth. Incorrect recognitions constitute False Positives (FP).
The evaluation set comprises frames in which a license plate was physically
present and visible; frames with no detectable plate are excluded.
Current text:
"Training of the YOLOv8s detection model was performed on 1,240 annotated frames from a separate collection session"
Problem: "A separate collection session" is vague. Where were these 1,240 frames collected? Same campus? Different campus? Public dataset? A reviewer will want to know if there is domain leakage between training and test data.
Suggested rewrite:
- Training of the YOLOv8s detection model was performed on 1,240 annotated
- frames from a separate collection session and validated on a held-out
- set of 310 frames
+ Training of the YOLOv8s detection model was performed on 1,240 annotated
+ frames collected at the same campus entry gate during a prior data
+ collection session conducted [X months] before the evaluation period.
+ Annotations were created using [Roboflow / LabelImg / CVAT] in YOLO
+ format with a single class (license_plate). A held-out validation set
+ of 310 frames (20% stratified split) was used for early stopping.
+ The 747-image OCR evaluation set was collected in a temporally disjoint
+ session with no vehicle overlap verified, ensuring no data leakage
+ between detection training and recognition evaluation.Important
Fill in the bracketed [X months] and [Roboflow / LabelImg / CVAT] with the actual values. If the training data came from Roboflow or a public dataset, state the exact source and license.
Problem: Every comparison in Table 1 and Table 3 is cross-dataset. You acknowledge this in the Table 1 caption and in the limitations section, but the abstract, contribution (vi), and conclusion still use language like "outperforming" and "surpassing" without qualification.
Suggested fix — add a disclaimer paragraph before Table 3:
Note on cross-dataset comparison. No shared public benchmark exists for
Indian ANPR evaluation. The accuracy figures for Lim et al. [3] and
Agrawal & Joshi [11] in Table 3 are reproduced from their respective
original evaluations on different private datasets with different camera
configurations, vehicle populations, and environmental conditions. The
numbers are presented to contextualise the proposed system's performance
relative to the best available reference points, but should not be
interpreted as controlled comparisons on a common test set. The ablation
study in Table 6 provides the only controlled comparison, as all
configurations are evaluated on the same 747-image dataset.
Problem: Table 7 reports Raspberry Pi 5 latency (520ms/frame, 1.9 FPS), but you never report latency on the development hardware (Apple Silicon M-series) where the 97.32% accuracy was actually measured. Reviewers comparing against prior methods will want to know the compute cost of your accuracy.
Suggested fix — add a Table 7b or expand Table 7:
Table 7: Per-stage latency comparison across deployment targets (mean ± std, ms).
Stage Apple M-series Raspberry Pi 5
─────────────────────────────────────────────────────────────────────────
Plate detection (YOLOv8s) 18 ± 4 145 ± 23
Preprocessing (CLAHE + threshold) 5 ± 2 32 ± 8
OCR ensemble (PaddleOCR + EasyOCR) 85 ± 15 340 ± 67
IndianPlateFormatter 1 ± 0.3 3 ± 1
End-to-end (single frame) 109 ± 18 520 ± 89
Effective throughput ~9 FPS 1.9 FPS
Tip
You should run a quick timing benchmark on your Mac to fill in these numbers. The time.perf_counter() instrumentation already exists in video_pipeline.py. Process 100 frames and report the mean.
Also consider adding a latency column to Table 3 where available:
Table 3: Overall recognition performance on CSD-747.
Method Acc.(%) CER(%) P R F1 Latency
Lim et al. [3] (Malaysian) 98.64 — — — — —
Agrawal & Joshi [11] (Indian, 3 st.) 95.82 — — — — —
Proposed (M-series dev hardware) 97.32 0.27 0.973 1.000 0.986 ~109 ms
Proposed (Raspberry Pi 5) 97.32 0.27 0.973 1.000 0.986 ~520 ms
Current limitations text already says:
"the comparison with prior work is cross-dataset as no shared public Indian ANPR benchmark exists"
This is good. However, it's buried as the last sentence of the limitations paragraph. Given how central this issue is (Points 6, 10, 11), it deserves more prominence.
Suggested rewrite of the full limitations paragraph:
Limitations. The principal limitation of this evaluation is the absence of
a shared public Indian ANPR benchmark, which means all comparisons with
prior work in Table 1 and Table 3 are cross-dataset and reflect different
camera configurations, vehicle populations, lighting conditions, and plate
diversity. The reported 1.50 percentage point improvement over the prior
best [11] should be interpreted in this context; establishing definitive
superiority would require evaluation on a common test set, which we
encourage the community to develop. The evaluation dataset (CSD-747) is
institution-specific and not publicly released due to privacy constraints,
further limiting direct reproducibility. The two-wheeler evaluation subset
(n=42, of which scooters n=11) is too small for statistically robust
sub-category conclusions. The 1.9 FPS throughput on Raspberry Pi 5 is
sufficient for campus gate surveillance but insufficient for high-speed
highway deployment without further OCR engine optimisation. The system
handles 38 Indian RTO state codes but does not extend to special
diplomatic or military plate formats.
Tip
Honest, prominent limitations actually help with reviewers. It shows you understand the work's scope. Hiding limitations in the last sentence signals lack of self-awareness, which is worse.
| Priority | Point | Effort | Impact on Acceptance |
|---|---|---|---|
| 🔴 Critical | 6 Cross-dataset "outperforming" claim | Low (rewording) | Prevents rejection |
| 🔴 Critical | 7 Hyperparameter justification | Medium (add sensitivity analysis or honest statement) | Prevents major revision |
| 🟡 Important | 1 Title length | Low (pick alternative) | First impression |
| 🟡 Important | 2 Gap–contribution mapping | Medium (restructure §1) | Strengthens narrative |
| 🟡 Important | 3 Two-wheeler statistical qualification | Low (add CI caveat) | Prevents statistical criticism |
| 🟡 Important | 9 Training data provenance | Low (1 sentence) | Prevents data leakage question |
| 🟡 Important | 10 Cross-dataset disclaimer before Table 3 | Low (add paragraph) | Consistency with Table 1 |
| 🟡 Important | 12 Prominent limitations | Low (rewrite paragraph) | Reviewer goodwill |
| 🟢 Nice | 4 Footnote cleanup | Low | Readability |
| 🟢 Nice | 5 Table heading length | Low | LNCS formatting |
| 🟢 Nice | 8 Metrics as bullet points | Low | Readability |
| 🟢 Nice | 11 Dev hardware latency | Medium (run benchmark) | Completeness |