A Lightweight Convolutional Neural Network for Real-Time Recognition of Hand-Drawn Geometric Shapes
cs.CV · cs.LG
Quick verdict
Compact 97.9k-param CNN hits 96.01% val accuracy on 4 hand-drawn shapes (2k images, one author); honest limits, but no test set or baselines and thin novelty.
Summary
This paper describes a self-contained pipeline for recognizing four hand-drawn geometric shapes (circle, square, rectangle, triangle). The author collected 2,000 hand-drawn 28×28 images — all drawn by a single contributor — and trained a compact CNN of 97,956 parameters: three convolutional blocks (16/32/64 filters) with max-pooling, embedded augmentation layers (horizontal flip, rotation, zoom), dropout, and a 128-unit dense head. Training used Adam (lr 0.001) with sparse categorical cross-entropy computed on logits. On an 80/20 train-validation split the model reaches 94.80% training accuracy, 96.01% validation accuracy, and 0.1437 validation loss. A Tkinter desktop app performs real-time inference from mouse drawings after grayscale, resize, and edge-enhancement preprocessing. The architecture was recovered from the released Keras checkpoint; code, trained model, and per-class Kaggle datasets are public. Notably, the paper candidly reports its own gaps: the per-class distribution and confusion matrix are unrecorded, epochs and batch size are unknown, and no independent test set exists. The comparison with Eitz et al. (56% machine accuracy on 250 sketch classes) and Alam et al. (98% on 8 classes, 20,000 images) is explicitly labeled illustrative. Main weaknesses: the headline accuracy may be optimistic given the absent test set and single-artist data; no classical or simple baselines are evaluated; and novelty over prior geometric-shape classifiers is modest.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 97,956-parameter CNN reaches 96.01% validation accuracy (0.1437 validation loss) on a 4-class hand-drawn shape task using only 2,000 self-collected 28×28 images.
- Validation accuracy exceeding training accuracy (96.01% vs 94.80%) is attributable to train-time augmentation and dropout regularization.
- A sub-100k-parameter CNN suffices for a minimal four-class geometric vocabulary without transfer learning, pretrained backbones, or hybrid ensemble classifiers.
- The complete pipeline — dataset, trained checkpoint, and Tkinter app — is publicly released, with the architecture recovered exactly from the released .h5 file.
Strengths
Unusually candid limitations reporting
Section VIII explicitly lists the missing test set, the unrecorded per-class distribution and confusion matrix, and the undocumented epochs/batch size, and states the confusion matrix 'is not fabricated here' — rare transparency that pre-empts most overclaiming.
Full artifact release enables verification
GitHub repo [1], the trained checkpoint (shapedetector_model_4b.h5), and four per-class Kaggle datasets [2]-[5] are public; Table I's architecture was recovered layer-by-layer from the checkpoint and its counts sum exactly to 97,956 (448+4,640+18,496+73,856+516).
Honestly scoped cross-study comparison
Technically sound design choices
Weaknesses
No independent test set
Single-contributor data with unknown class balance
No baselines or ablations on own data
Training configuration gaps block exact replication
Thin novelty and a related-work gap
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper documents an end-to-end desktop system for classifying four hand-drawn geometric primitives (circle, square, rectangle, triangle). Its claims are modest and mostly well-scoped: (1) a 97,956-parameter CNN (Table I) reaches 94.80% training / 96.01% validation accuracy with 0.1437 validation loss on a 2,000-image self-collected 28×28 dataset (Table II); (2) validation accuracy exceeding training accuracy is attributed to train-time augmentation and dropout (Sec. VII) — a technically correct explanation; (3) a sub-100k-parameter model suffices for a four-class geometric vocabulary without transfer learning (Sec. VII). Sec. VIII pre-empts most obvious overreach, and Table III explicitly disclaims ranked comparison. The one assertion inviting pushback is the claim that 'no public dataset of hand-drawn geometric shapes suited to this task was readily available' (Sec. III-A): Google's Quick, Draw! (2016+) contains raster sketches of exactly these four categories at massive scale, and the related-work section never engages with it.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 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 this hand-drawn shape recognition model achieve?
The paper reports 94.80% training accuracy and 96.01% validation accuracy with a 0.1437 validation loss on a 4-class problem (circle, square, rectangle, triangle), evaluated on a 400-image validation split of 2,000 images. There is no independent test set, so accuracy on new users' drawings may be lower.
Is the dataset and code for this paper publicly available?
Yes. The 2,000 images are split into four per-class Kaggle datasets (triangles, squares, rectangles, circles), and the source code plus the trained Keras model file are on the author's GitHub repository (Handwritten-Geometric-Shape-Detector). However, training epochs and batch size were never recorded.
How big is the neural network, and can it run in real time on a normal computer?
The CNN has 97,956 trainable parameters across three convolutional blocks (16, 32, 64 filters) and a 128-unit dense layer. The paper states it supports real-time, CPU-only inference inside a Tkinter desktop application where users draw shapes with the mouse.
Can this model recognize shapes drawn by other people?
Unknown. All 2,000 training images were drawn by the paper's single author, and no cross-user evaluation was performed. The paper itself flags single-contributor handwriting as a key limitation that may limit generalization to different drawing styles, stroke widths, and input devices.
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.
- Figures 1 and 2 are present only as captions/placeholders in the extracted text, so the pipeline diagram and accuracy plot could not be inspected.
- The arXiv margin date is Sep 2026 while Table III dates the underlying work to 2021; this review treats it as a late writeup of an older project.
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.