CNN portfolio project 04

100-class image classification, directly in your browser.

A ResNet50 transfer-learning model trained on CIFAR-100, converted into TensorFlow.js, and served as a static GitHub Pages application. Images stay on this device.

Try the classifier View model details

67.09%Test accuracy

91.2%Top-5 accuracy

100Fine classes

32×32 RGB↓Resize + preprocess↓ResNet50 backbone↓Pooling + dense head↓100-class softmax

Interactive inference

Classify an image

Waiting for TensorFlow.js…

1. Select an image

Use PNG, JPEG, WebP, or BMP. For privacy, the file is processed locally and is not uploaded.

Choose or drop an image Maximum recommended size: 10 MB

Generated test images:

2. Preview

Your image preview will appear here.

3. Prediction

Run inference to view the predicted class, confidence, and top probabilities.

Technical design

Transfer learning with residual features

ResNet skip connections help information and gradients move through deep networks. This project reuses ImageNet visual features, freezes the backbone during the supplied experiment, and trains a CIFAR-100 classification head.

BackboneResNet50

Browser input96 × 96 × 3

Output100 probabilities

RuntimeTensorFlow.js

Responsible use

Interpret predictions carefully

This educational demo may classify images incorrectly, especially images that are unclear, low-quality, out of distribution, or unlike CIFAR-100. Do not use it as the sole basis for medical, legal, safety-critical, security, hiring, insurance, financial, or production decisions. Do not upload private, sensitive, confidential, copyrighted, or personally identifiable images.