https://github.com/bhavyemathur/defect-detection

Machine Learning for unsupervised & single-shot quality inspection & defect-detection on the assembly line.
https://github.com/bhavyemathur/defect-detection

defect-detection machine-learning single-shot-detection unsupervised-learning

Last synced: 7 months ago
JSON representation

Machine Learning for unsupervised & single-shot quality inspection & defect-detection on the assembly line.

Host: GitHub
URL: https://github.com/bhavyemathur/defect-detection
Owner: BhavyeMathur
Created: 2024-06-28T08:21:57.000Z (almost 2 years ago)
Default Branch: main
Last Pushed: 2024-08-07T15:01:09.000Z (almost 2 years ago)
Last Synced: 2024-08-07T19:15:18.187Z (almost 2 years ago)
Topics: defect-detection, machine-learning, single-shot-detection, unsupervised-learning
Language: Jupyter Notebook
Homepage:
Size: 1.09 GB
Stars: 0
Watchers: 1
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md

Awesome Lists containing this project

README

This repository contains code developed at [AIVolved](https://aivolved.in/) for quality testing using Machine Learning & Artificial Intelligence.
Several goals were pursued on key datasets including:

1. Eye-patch shift detection using non-ML computer vision techniques (92.98% accuracy)
2. Clustering & SVD on ResNet output for unsupervised defect detection in soap (99.55% acccuracy)
3. Unsupervised, single-shot defect detection in soap using the Fourier Transform (100% accuracy)
4. Cut detection in shampoo using Sobel & Canny filters + Hough Transform (N/A)

# Eye-patch shift Dataset

[eyeshift.ipynb](eyeshift.ipynb) contains code that identifies defects in eye-patches for shampoo packets.
First, a YOLOv8 model identifies horizontal and vertical cuts, then a linear regression is performed through the horizontal cuts,
and eye-patches outside a threshold are categorised as defective.

**Accuracy:** 92.98%

![img.png](assets/eyeshift.png)
![img.png](assets/eyeshift2.png)

# Soap Dataset

[soap-binary-classifier.ipynb](soap-binary-classifier.ipynb) uses a simple fully-connected layer on the outputs from ResNet18,
fine-tuned on a dataset of soap to classify as either defective or non-defective.

**Accuracy:** 100%

Non-Defective
Defective

[soap-feature-clustering.ipynb](soap-feature-clustering.ipynb) is an unsupervised approach to defect detection in
this dataset where features from the ResNet18 output undergo a Singular Value Decomposition (SVD) and are then clustered using Birch.

**Accuracy:** 99.55%

Ground Truth
SVD & Clustering
Prediction

[soap-autoencoder.ipynb](soap-autoencoder.ipynb) is actually a U-net which attempts to reconstruct masked images of soap to predict defective pieces by correcting errors.

**Accuracy:** untested.

Masked Input
Prediction

[soap-fourier-analysis.ipynb](soap-fourier-analysis.ipynb) is a single-shot, unsupervised method for defect detection on a normalised dataset.
A non-defective single-shot reference image is chosen and the squared complex-difference between its Fourier Transform and all other images in the dataset are compared and clustered.

**Accuracy:** 100%

Fourier Transform of Soap
Histogram of Differences to Reference

# Shampoo Dataset

[shampoo.ipynb](shampoo.ipynb) contains code that identifies defective cuts in shampoo packets.
First, vertical cuts are extracted using a YOLOv8 model, the cuts are equalised and normalised and a sobel filter is applied to enhance edges.
A Canny edge-detector is used followed by a Hough transform to identify cuts.

**Accuracy:** visually excellent. No quantitative measure.

Masked Input
Cuts Extracted from YOLOv8