🏭 Industrial Defect Generator 🏭

Description

This library aims to give a simple access to the proposed defect generation approaches for industrial defects.

Access to demo

A demo is available in HuggingFace : Link

Base usage

The library is designed to be straightforward to use with a single function that generate one or several defects and masks

image = Image.open('path/to/your/image').convert('RGB')
defGen=DefectGenerator(image.size)
defect,msk=defGen.genDefect(image,defectType=["nsa"])

Arguments

• image (Pil Image) : The input image
• defectType (List[str]) : A list of string containing the defect types, candidates are "nsa", "structural", "textural", "blurred", "cutpaste", "cutpasteScar"
• category (str, default="") : The object category to mask the background if necessary
• return_list (bool, default=False) : To generate the output as a list of defective image and masks instead of a single defective image and mask

Output

If return_list is False :

• imageDefect (torch.Tensor) : The generated defective image
• mask (torch.Tensor) : The mask of the generated defect

If return_list is True :

• imageDefect (List[torch.Tensor]) : The list of the generated defective images
• mask (List[torch.Tensor]) : The list of the masks of the generated defects

Notes

• If defectType contains multiple elements and return_list is false, the function generate a random defect among the possibilites in defectType. If return_list is True, the function return a list of images and a list of masks with each defects from defectType.
• To modify in details the defect generation, you may want to modify the files in source/

Implemented methods

Methods currently implemented in the library

NSA

Article : https://arxiv.org/pdf/2109.15222.pdf
Code inspiration : https://github.com/hmsch/natural-synthetic-anomalies

DRAEM

Article : https://arxiv.org/pdf/2108.07610.pdf
Code inspiration : https://github.com/VitjanZ/DRAEM

MemSeg

Article : https://arxiv.org/ftp/arxiv/papers/2205/2205.00908.pdf
Code inspiration : https://github.com/TooTouch/MemSeg

CSE

Article : https://arxiv.org/pdf/2403.01859.pdf
Code inspiration : https://github.com/SimonThomine/CSE

CutPaste

Article : https://arxiv.org/pdf/2104.04015.pdf
Code inspiration : https://github.com/LilitYolyan/CutPaste