putputloh / Surface-Defect-Detection

Leverage the power of Deep Learning to automatically detect, segment and classify surface defects 🤯🤯

Surface Defect Detection

Introduction:

🙅‍♂️ Human quality inspection
  ➣ Some defects tend to be miniature and difficult to identify
  ➣ Various appearances and characteristics of defects often lead to misclassification of their categories
🙆 AI inspection using deep learning semantic segmentation models
  ➣ Segment and classify defects in the input image by predicting pixel-labelled segmentation masks

Objectives:

  ➣ Research and develop deep learning semantic segmentation models for automatic defect detection
  ➣ Evaluate and compare CNN-based, Transformer-based and hybrid CNN-Transformer models in various aspects
  ➣ Explore and implement advanced weakly supervised semantic segmentation algorithms / frameworks

Datasets

  ➣ Kolektor Surface-Defect Dataset 2 (KolektorSDD2/KSDD2)
  ➣ Magnetic Tile Surface Defects

Results:

RANKINGS
Fully-Supervised	mIoU		Weakly-Supervised	mIoU
	Magnetic Tile	KolektorSDD2		Magnetic Tile	KolektorSDD2
🥇TransDAU-Net	83.58	79.36	🥇SEAM	-	67.00
🥈TransU-Net	83.19	78.92	😩CAM	-	39.84
🥉Double U-Net	82.81	78.43
🏅U-Net	80.65	78.21
😩SETR	74.25	72.98
😩FCN	66.95	-
😩DilatedNet	62.64	-

** Highlighted models = Self-designed models

Sample Prediction Masks:

Fully-Supervised	Magnetic Tile
	KolektorSDD2
Weakly-Supervised	KolektorSDD2