Visual Regression Testing with Neural Networks
summary
Convolutional autoencoder reproduces 256×256 grayscale screenshots with an 8-neuron bottleneck and flags pages that the network cannot reconstruct well so differences indicate visual regression. Implementation runs in Colab + TensorFlow on GPU, compressing images to eight values then decoding them; training on a golden master reaches ~99% reconstruction on training while unknown or broken pages show significantly lower scores. Proof-of-concept catches overflows and missing buttons; thresholding controls inspection volume.
AI-written summary · about
chapters (8)
- 0:00Introduction to the topic
- 2:44Regression testing challenges
- 11:04Machine learning basics
- 17:10Autoencoders for testing
- 21:42Technical implementation
- 27:35Training and validation
- 33:58Limitations and conclusions
- 39:41Q&A session
related talks
34mTesting Live Video Streams At Scale
01h 27mHands-On Computer Vision with PyTorch
42mDeep Learning for Computer Vision
04mUsing Computer Vision to Understand Big Visual Data
28mVisual Search
45mAnomaly detection with TensorFlow
47mDeep Learning in Computer Vision
34mNeural networks: insane in the membrane