Deep Transfer Learning for Computer Vision
summary
Nanoscale semiconductor defect pipelines combine fine-tuned ResNet50 deep features with traditional descriptors like Zernike moments and GLCM, fuse them into a classifier stack that reaches about 95% accuracy across four defect classes, and uses a Faster R-CNN backend to localize and count defects from SEM/AFM imagery. Image augmentation handles tiny datasets, a clustering UI (hierarchical + affinity propagation) surfaces repeatable clusters and singleton outliers, and the system is deployed for recipe debugging.
AI-written summary · about
chapters (10)
- 0:00Intro
- 1:15Defect Classification - Key Objectives
- 2:22Defect Classification - Types of Defects
- 2:36Automated Defect Classification - Architecture
- 3:47ADC - Hybrid Classifiers
- 4:42ADC - Evaluation Benchmarking
- 5:10Defect Clustering - Business Objective
- 5:41Automated Defect Clustering - Architecture
- 6:38ADC - Hybrid Feature Extractor
- 6:54ADC - Agglomerative Hierarchical Clustering
related talks
52mTransfer Learning and Finetuning Deep Convolution Neural Network
59mDeep Neural Networks for Video Applications
55mDeep Learning in Medical Image Diagnostics
01h 30mDeep Learning Techniques for Monocular Depth Estimation
24mImage Recognition for safety on the factory floor
41mPractical Aspects of Applying Modern Computer Vision Technologies
47mDeep Learning in Computer Vision
12mCascade models in computer vision to boost accuracy and performance