VOZAP

Solution pipeline
Given data
Synthetic data
Labeling
Neural Network for classification task
Neural Network for segmentation task
Classification task
Achievements meet numbers
Tools

Intro

Our challenge is all about Semiconductor microscopy process, Image processing, and Machine Learning. We started with 111 given input images (increased to 1500+ with our synthetic dataset) and developed a tool that can do classification and segmentation tasks. The result of our processing toolchain is output for MeRit System.

What it does

Classification
Segmentation

How we built it

Data: augment, synthetic
Preprocessing
Classification
Semantic segmentation
Scalable interaction with a user (or API)

Challenges we ran into

Photomasking challenge (#2): classification and segmentation

-- For classification have used pre-trained vgg-11 and Resnet-18 with the change in final layer for our three class classification in photoMask. On the given training set we have achieved 94 % (vgg-11) on validation set and on given hold-out test set we achieved 82 % accuracy.

-- For Segmentation, we have used Mask-RCNN on top of our classification achieved 80 % correct masking on a given samples

--- For Eclipse, we used the least square optimization to get the correct shape based on the polygonal mask return from Mask-RCNN

Accomplishments that we're proud of

Our unique dataset
Cloud environment for NN training

What we learned

Data labeling is hard
MS Azure is a useful tool

What's next for VOZAP

Improve an accuracy
Publish the dataset

Github

https://github.com/rimidal/vozap-team

Classification results

Using cpu {'BORDERTOUCHER': 0, 'COMPLETE_FEATURE': 1, 'NO FEATURE': 2} 0 BORDERTOUCHER/BD01.tif 1 BORDERTOUCHER/BD02.tif 1 BORDERTOUCHER/BD03.tif 1 BORDERTOUCHER/BD04.tif 1 BORDERTOUCHER/BD05.tif 2 BORDERTOUCHER/BD06.tif 0 BORDERTOUCHER/BD07.tif 0 BORDERTOUCHER/BD08.tif 0 BORDERTOUCHER/BD09.tif 1 BORDERTOUCHER/BD10.tif 1 COMPLETE_FEATURE/OK01_1.tif 1 COMPLETE_FEATURE/OK01_2.tif 1 COMPLETE_FEATURE/OK01_3.tif 1 COMPLETE_FEATURE/OK01_4.tif 1 COMPLETE_FEATURE/OK02_1.tif 1 COMPLETE_FEATURE/OK02_2.tif 1 COMPLETE_FEATURE/OK02_3.tif 1 COMPLETE_FEATURE/OK02_4.tif 1 COMPLETE_FEATURE/OK03_1.tif 1 COMPLETE_FEATURE/OK03_2.tif 1 COMPLETE_FEATURE/OK03_3.tif 1 COMPLETE_FEATURE/OK03_4.tif 1 COMPLETE_FEATURE/OK04_1.tif 1 COMPLETE_FEATURE/OK04_2.tif 1 COMPLETE_FEATURE/OK04_3.tif 1 COMPLETE_FEATURE/OK04_4.tif 1 COMPLETE_FEATURE/OK05_1.tif 1 COMPLETE_FEATURE/OK05_2.tif 1 COMPLETE_FEATURE/OK05_3.tif 1 COMPLETE_FEATURE/OK05_4.tif 1 COMPLETE_FEATURE/OK06_1.tif 1 COMPLETE_FEATURE/OK06_2.tif 1 COMPLETE_FEATURE/OK06_3.tif 1 COMPLETE_FEATURE/OK06_4.tif 1 COMPLETE_FEATURE/OK07_1.tif 1 COMPLETE_FEATURE/OK07_2.tif 1 COMPLETE_FEATURE/OK07_3.tif 1 COMPLETE_FEATURE/OK07_4.tif 1 COMPLETE_FEATURE/OK08_1.tif 1 COMPLETE_FEATURE/OK08_2.tif 1 COMPLETE_FEATURE/OK08_3.tif 1 COMPLETE_FEATURE/OK08_4.tif 1 COMPLETE_FEATURE/OK09_1.tif 1 COMPLETE_FEATURE/OK09_2.tif 1 COMPLETE_FEATURE/OK09_3.tif 1 COMPLETE_FEATURE/OK09_4.tif 1 COMPLETE_FEATURE/OK10_1.tif 1 COMPLETE_FEATURE/OK10_2.tif 1 COMPLETE_FEATURE/OK10_3.tif 1 COMPLETE_FEATURE/OK10_4.tif 0 COMPLETE_FEATURE/OK11_1.tif 1 COMPLETE_FEATURE/OK11_2.tif 1 COMPLETE_FEATURE/OK11_3.tif 1 COMPLETE_FEATURE/OK11_4.tif 0 COMPLETE_FEATURE/OK12_1.tif 1 COMPLETE_FEATURE/OK12_2.tif 1 COMPLETE_FEATURE/OK12_3.tif 1 COMPLETE_FEATURE/OK12_4.tif 2 NO FEATURE/N01_1.tif 2 NO FEATURE/N01_2.tif 1 NO FEATURE/N02_1.tif 1 NO FEATURE/N02_2.tif 0 NO FEATURE/N03.tif 1 NO FEATURE/N04.tif 0.8125