Inspection using ML

Antonio Grimaldo

Why?

• Machine learning algorithms are not common
• Traditional methods require a lot more time to develop a solution and more experience
• Traditional methods don't work well with textures or patterns

The Challenge

Machine vision lab

Chocolates

The challenge

The dataset consisted of 50 images:

• 41 images of "non-defective" chocolate and  9 images of "defective" chocolate

How it works?

Image

Model

Output

SVM

-1 or 1

Chocolate

techniques

1. One class SVM
2. One class SVM using PCA
3. One class SVM with cropped images
4. One class SVM with cropped images and PCA
5. Isolation trees with cropped images (Bonus)
6. One class SVM with cropped and filtered images (Bonus)

Isolation Forest            OneClass SVM

Support Vector MAchines

How can we separate this?

 High dimensionality!!

Principal Component Analysis

• It is a procedure for identifying a smaller number of uncorrelated variables, called "principal components", from a large set of data.
• It reduces the dimensionality of data 

Results

Filtered Image

Average of all images

One Class SVM with cropped images

Accuracy: 98%

FiN.