Image Robustness Demo
Upload any image and use the sliders to apply image degradations — blur, noise, JPEG compression, brightness, and contrast shifts — while watching MobileNetV2's top predictions and confidence scores update in real time. This directly illustrates one of the core themes from my PhD research: how image quality variation affects deep neural network performance.
Loading MobileNetV2…
Original
Running inference…
Degraded
Adjust the sliders below.
Degradation Controls
Research context: This demo connects directly to my work on image quality and model robustness. My quality filtering post covers how quality mismatch between training and deployment degrades classifier performance — and why filtering the lowest-quality training data often beats architectural changes. The model here is MobileNetV2 trained on ImageNet, running entirely in your browser via TensorFlow.js — no data leaves your machine.