CNN image classification viva — layers, metrics, and failure modes

CNN image classification projects win panels when you sound like you trained and tested a model — not when you recite “convolution extracts features” from a memorized Medium post. Indian B.Tech examiners have seen plant disease, traffic signs, handwritten digits, and medical X-ray clones for years. They separate students who can explain a block diagram, a confusion matrix, and a wrong prediction from students who only click Predict on a Streamlit button.

This guide is a viva playbook for kits such as Plant Disease Classification and Traffic Sign Recognition. Pair it with three patterns for AI projects and Streamlit demo habits. Face-based systems (attendance) share some CV questions but different privacy angles — see Face Recognition Attendance only if that is your topic.

Opening pitch (40 seconds)

We built an image classification system using a convolutional neural network. Users upload an image; the model returns a class label and confidence scores. We trained on a labeled dataset with train/validation/test splits, monitored accuracy and loss, and evaluated with a confusion matrix. The demo shows correct predictions and at least one failure case. Limitations include domain shift, class imbalance, and CPU inference latency. We are not claiming medical or safety certification.

Say it without looking at slides. Then invite the first question.

What “CNN” must mean in your mouth

A CNN stacks learnable filters that detect local patterns (edges → textures → parts → object-ish concepts), with pooling to reduce spatial size, and dense layers (or global pooling + softmax) for class probabilities. You do not need to derive backprop on the whiteboard unless asked. You do need:

1. ▸Input shape (e.g. 224×224×3).
2. ▸Rough depth (how many conv blocks).
3. ▸Number of output classes.
4. ▸Loss (categorical/sparse categorical cross-entropy).
5. ▸Optimizer (Adam is common — say why you kept defaults or what you tuned).

Transfer learning honesty

If you used MobileNet/EfficientNet/VGG pretrained on ImageNet and fine-tuned the head, say so proudly. Transfer learning is an engineering decision, not cheating. Claim “we designed a novel architecture” only if you did. Panels punish novelty theater.

If you trained a small custom CNN from scratch on a tiny dataset, defend that as a teaching model and list accuracy limitations honestly.

Dataset story — where most vivas are won

Be ready for: source, number of images, number of classes, train/val/test ratios, augmentation, and license/ethics.

Example plant disease narrative: public leaf dataset (name it), N classes (healthy + diseases), stratified split 70/15/15, augment with flip/rotation/brightness, resized to model input. Reject “we downloaded images from Google” without class balance talk.

Class imbalance

If one class dominates, accuracy lies. Mention precision/recall/F1 for minority classes or weighted metrics. If you only report accuracy, practice the sentence: “Accuracy is X; minority class recall is lower at Y — see confusion matrix.”

Data leakage

Same leaf photographed twice in both train and test inflates scores. Augmentation copies are not independent tests. Saying you split by original file IDs before augmenting sounds mature.

Metrics you should be able to define in one line

Accuracy: correct / total.
Precision: of predicted positives, how many truly positive.
Recall: of actual positives, how many found.
F1: harmonic mean of precision and recall.
Confusion matrix: rows true, columns predicted (confirm your library’s orientation).
Top-k: useful when UI shows top-3 diseases.

For multi-class plant disease, macro vs micro F1 may appear in reports — know which you printed.

Overfitting signals

Train accuracy high, val accuracy flat or dropping; train loss ↓ val loss ↑. Remedies you tried: dropout, augmentation, early stopping, fewer dense units, transfer learning freeze/unfreeze schedule. Name one remedy you actually used.

Architecture whiteboard mini-script

“Input image → preprocessing normalize → Conv+ReLU+Pool blocks extract features → flatten or global average pool → dense dropout → softmax over C classes → argmax for label, probabilities for confidence UI.”

If asked “what does the first layer learn?”, say low-level edges/colors; deeper layers combine them. Avoid mystical “the AI understands leaves.”

Why not a plain MLP on pixels?

CNNs share weights spatially and respect locality; full dense-on-pixels explodes parameters and ignores neighborhood structure. That contrast is a classic examiner question.

Demo path that impresses

1. ▸Show sample gallery image → correct class + confidence.
2. ▸Upload a slightly harder image → still correct or show top-2.
3. ▸Upload a failure (blurry, wrong crop, out-of-domain object) → wrong or low confidence — narrate why.
4. ▸Show training curves or confusion matrix screenshot from report/results folder.
5. ▸Optional: latency note (“~X ms on CPU”).

Never demo only green-path images curated to 99%. See what examiners look for in a demo.

Confidence is not calibrated truth

A model can be 0.92 confident and wrong. Say confidence is a softmax score, not a guarantee. Thresholding “don’t show prediction if max prob < 0.5” is a good product sentence if implemented.

Failure modes — memorize five

1. ▸Domain shift: lab photos vs field photos; phone cameras differ.
2. ▸Occlusion / framing: leaf not centered; multiple leaves.
3. ▸Lighting: heavy shadow, yellow indoor light.
4. ▸Similar classes: early vs late blight confusion.
5. ▸Out of vocabulary: random cat photo forced into disease classes — need “unknown” strategy or low confidence messaging.

For traffic signs: motion blur, night scenes, partial signs, country-specific sign sets not in training.

Training practicalities for student laptops

Training on Colab/Kaggle GPU, inference on laptop CPU is a normal story. Save .h5 / SavedModel / .keras weights in the repo or drive link your guide approves. Document library versions (TensorFlow/Keras/OpenCV). If Streamlit cloud is flaky at venue, run local.

Reproducibility

Fixed seeds help but GPUs are not perfectly deterministic — do not overclaim bit-identical training. Version the dataset folder hash or download script.

Streamlit UI talking points

Upload widget, predict button, class label, probability bar chart, example images, optional Grad-CAM if you truly implemented it. Empty Grad-CAM buttons are viva traps — remove or finish.

Do not bury the brand of your college project under six metric cards on the first screen; one clear prediction panel is enough for viva.

Report chapter mapping for CNN kits

Chapter 1: manual diagnosis pain / road safety motivation. Chapter 2: brief CNN literature (cite 2–3 papers/books, not twenty unread links). Chapter 3: requirements, non-goals (not a replacement for agronomists/doctors). Chapter 4: architecture diagram, dataset stats. Chapter 5: training pipeline, hyperparameters table. Chapter 6: metrics, confusion matrix, error analysis. Chapter 7: limitations and future work (more data, mobile deploy, unknown class).

Extended viva Q&A

Q: Softmax vs sigmoid? A: Softmax for mutually exclusive multi-class; sigmoid for multi-label. Know which your problem is.

Q: What is a kernel/filter? A: Small weight matrix slid over feature maps producing activations.

Q: Padding? Stride? A: Padding preserves size; stride controls downsampling. Give your settings if known.

Q: Batch normalization? A: Stabilizes training by normalizing layer inputs; mention only if used.

Q: Why ReLU? A: Sparse activation, mitigates vanishing gradients vs sigmoid in deep stacks — keep it short.

Q: Can this run on mobile? A: Possible with TFLite/ONNX future work; current demo is PC/Streamlit.

Q: Ethics? A: Plant disease advice is assistive; medical claims need disclaimers. Traffic systems are experimental, not installed on vehicles.

Q: How is this ML not “just if-else”? A: Parameters learned from data minimize loss; rules were not hand-coded per class texture.

Differentiating when half the batch picked “plant disease”

Change crop set, add a confidence threshold UX, write error analysis on 20 hard images, compare two backbones (MobileNet vs custom) with a table, or add a simple farmer-language suggestion text per class. See same project differentiate. Cosmetic theme color alone is weak.

Common mistakes

Saying “100% accurate.” Hiding the confusion matrix. Unable to state number of classes. Claiming real-time video when you classify single uploads. Reading Keras summary line-by-line without explaining. Blaming the kit for a class you cannot define. Using test images that were in training. No failure demo. Calling Transformers/ViT when the repo is a CNN.

Team viva splits for CNN projects

One member owns dataset + augmentation story, one owns model + training curves, one owns Streamlit + demo failure case. Everyone must still summarize end-to-end. “I only did UI” fails when asked how softmax produces the label on screen.

Pre-viva checklist

• ▸[ ] Model loads offline without surprise downloads mid-viva
• ▸[ ] Three demo images ready (easy, medium, fail)
• ▸[ ] Confusion matrix accessible (slide or results folder)
• ▸[ ] Hyperparameter table memorized (epochs, batch, LR, input size)
• ▸[ ] Class list printable in under ten seconds
• ▸[ ] Disclaimer sentence for domain (agri/medical/traffic)
• ▸[ ] Backup screencast if webcam/projector fights Streamlit

Connecting to other AI patterns on FinalYearKit

Classification ≠ RAG ≠ classical ML dashboards. If a panel asks why you did not use ChatGPT vision, answer with cost, offline needs, and controlled label space — your CNN predicts among known classes with a fixed head. RAG kits solve grounded Q&A over documents; different problem. Keep boundaries crisp via three patterns for AI projects.

Hyperparameter table — practice saying numbers

Examiners often ask for batch size, epochs, learning rate, and input resolution in one breath. Example honest line: “Batch 32, 25 epochs with early stopping patience 5, Adam at 1e-4, images 224×224, freeze base for 5 epochs then unfreeze last blocks.” Your numbers must match Chapter 5. If you changed them during a late retrain, update the report before external viva so slides, report, and speech agree.

Early stopping as a maturity signal

Saying you watched val loss and stopped when it plateaued beats “we always train 100 epochs.” If TensorBoard or CSV logger plots exist, keep a PNG in results/ for the projector.

Preprocessing pipeline (say it as a sequence)

Read image → convert color space if needed (BGR/RGB footguns with OpenCV) → resize → normalize (0–1 or ImageNet mean/std for transfer models) → expand batch dimension → model.predict. Mismatch between training preprocessing and demo preprocessing is a top cause of “it worked in Colab.” Verify both paths use the same function.

Augmentation — what to claim

Horizontal flip, small rotation, zoom, brightness — common and defensible for leaves and signs. Do not claim MixUp/CutMix unless coded. Caution: vertical flip may be unrealistic for traffic signs; domain knowledge matters. Document augmentations in a bullet list in Chapter 5.

Evaluation protocol clarity

Single hold-out test set is fine for most colleges. K-fold is stronger but heavier — only describe it if you ran it. Always state whether the test set was touched during tuning. Ideal story: tune on validation, report once on test.

Edge deployment future work (without overpromising)

TFLite conversion, quantized weights, Android wrapper — list as future work with one sentence each. Do not show a phone mockup that is not connected. Panels have seen too many “future mobile app” slides with zero artifacts.

How to recover when prediction is hilariously wrong live

Stay calm. “This looks like domain shift / poor framing; confidence is Z; our error analysis includes similar cases.” Switch to a prepared gallery image, then return to analysis. Never accuse the examiner’s sample unfairly; thank them for a good stress test.

Explaining parameters vs hyperparameters

Parameters are learned weights. Hyperparameters are choices you set (learning rate, batch size, number of filters, dropout rate). Panels sometimes mix the terms; correcting gently with definitions shows command. If you tuned only learning rate and epochs, say that — do not invent a Bayesian optimization saga.

Parameter count as optional color

“MobileNetV2 has millions of parameters but is mobile-friendly due to depthwise separable convolutions” is a strong sentence if you used that backbone. For a tiny custom CNN, estimate roughly from your model.summary() and admit it is a teaching network.

Validation strategy when dataset is small

If you only have a few hundred images per class, say so and discuss risk of overfitting. Techniques: heavy augmentation, transfer learning, simpler heads, and cautious claims. Do not hide small-N behind three decimal accuracy places. Qualitative error analysis on twenty images can impress more than a brittle 99% claim.

Comparing classification to detection

Object detection returns boxes; classification returns one label (or multi-label tags) for the whole image. If your UI crops the leaf first, say classification assumes a single primary object. Traffic sign projects sometimes blur this line — clarify whether you classify a cropped sign image or detect signs in a full road frame. Scope honesty prevents “where is YOLO?” traps unless YOLO is in the repo.

Hardware narrative for Indian college labs

Many labs have no GPU. Your story should be: train on Colab/Kaggle, export weights, infer on CPU in the lab. Measure a rough latency once (time.time() around predict) so you can say “about X seconds per image on this laptop.” Fans spinning loudly during demo are normal; apologize once and continue.

Slide hygiene for CNN decks

One architecture diagram, one metrics table, one confusion matrix, one failure collage — enough. Avoid dumping twelve indistinguishable training-curve screenshots. Label axes. If the matrix is unreadable on a projector, print a handout or zoom a 3-class subset for discussion.

Related reading

See three patterns for AI projects, Streamlit final-year AI demos, and common viva mistakes.

Project kits

• ▸Plant Disease Classification — CNN + Streamlit leaf disease demo with confidence and gallery.
• ▸Traffic Sign Recognition — classic multi-class CNN theme with strong visual viva potential.
• ▸Face Recognition Attendance — related CV kit when panels compare classification vs recognition embeddings.

Takeaway: Win CNN vivas with a crisp pipeline story, real dataset and metric literacy, and a deliberate failure demo — not with buzzwords or perfect-accuracy claims. If your confusion matrix and preprocessing match the live app, you are already ahead of most of the batch.