2022 Samsung AI Challenge (3D Metrology) 1st place Solution

AI solution that produces depth image from SEM image

Presentation : Samsung-sem-1st.pdf

Solution Overview

Team member

정재윤, 위성진, 장준보

Code Running Environment

run at local

use docker image of paperspace gradient Scratch docker https://hub.docker.com/r/paperspace/gradient-base

docker tag
pt112-tf29-jax0314-py39-20220803

Machine Environment

Intel i9-10900 CPU, RTX3090 GPU, window 10 anaconda3 jupyter notebook

File Structure

┖ figures
  ┖ ~
┖ [ST]cycle_wider_BNGN.ipynb
┖ [CNN]Train.ipynb
┖ [INF]Inference.ipynb
┖ open.zip
┖ Readme.md
┖ Samsung-sem-1st.pdf

cycle folder

• model pth file and training image folder. It will be generated after [ST] or download weight.

[ST]cycle_wider_BNGN.ipynb

• Train GAN model for transfer simulation sem image to train sem image style.

[CNN]Train.ipynb

• Train efficientnet_b0 for Case Classifier

[INF]Inference.ipynb

• similarity model for getting depth image

open.zip

• data download from 2022 Samsung AI Challenge (3D Metrology)

Samsung-sem-1st.pdf

• code presentation written by korean

Dataset

We train and evaluate our model using the dataset from 2022 Samsung AI Challenge (3D Metrology)

we assume that you have downloaded it and placed based on File Structure.

Submission Process

Train Process

1. Run All code in [ST]cycle_wider_BNGN.ipynb for train domain transfer GAN and make domain transferred image that transferred from simulation sem to train sem. the transferred image will be saved in the ./processed_data folder. you need to log in wandb for tracking model progress

2. Run All code in [CNN]Train.ipynb for train case classification model. The classification model's pth file will be saved as ./cnn_classifier.pth.

---

Test Process

1. Run All code in [INF]Inference.ipynb to generate depth data from the test sem image. Test sem image and a zipped file will be saved in the ./Submission_F folder.