Industrial Vision · Surface Inspection
Classify a defect, retrieve visually similar samples
Drop a steel-surface image. A 6-class CNN returns the defect type with per-class probabilities; an autoencoder encoder runs L2-nearest-neighbour over a 1,800-image index and surfaces visually similar past defects.
Input
No image selected
Pipeline
ClassifierKeras CNN · 200²
EncoderConvAutoencoder_v2 · 512²
Latent dim65,536
IndexL2 / brute-force
Inference—
Awaiting input
No analysis yet
Upload an image to get the predicted defect class with full softmax, plus the k visually-closest past defects from the indexed reference set.
Predicted class
—
—
—
Visually similar · CBIR
—
Distance is L2 in the 65,536-dim encoder feature space. Lower = more similar.