You are on page 1of 7

Transfer Learning in Image Classification

Dr. Mohamed Waleed Fakhr


1

2/1/2021
2 3.1 Best models for image classification
in Covid-19 challenges
 Pre-trained CNN models (VGG, ResNet, DenseNet, InceptionNet, etc.):
1. Find tune the deep models after adding fully connected layers and training
with the new data
2. Use pre-trained models as feature extraction and apply traditional ML
techniques (SVM, RF and XGBoost)
 Train CNN architectures from scratch
 Beyond Challenges: More than 100 papers already available using deep
learning for classifying X-ray and CT-scan images
3.2. Examples of X-ray images used in challenges

Example-1: a) normal, b) COVID-19, and c) SARS chest X-ray images

Example-2: a) COVID-19, b)
Pneumonia, c) TB d) normal
chest X-ray images
4 3.3.1 Pre-Trained Deep Models
Have been used with 2 different approaches:
1- Use extra output layers and fine tune the architecture to the new data
(transfer learning)
5
3.3.2 Pre-Trained Deep Models

2- Using pre-trained deep neural network models as feature extractors


6
3.3.3 Training Deep Models from Scratch
Training known deep architectures from scratch:
 Logic: Most pre-trained models are trained on ImageNet, CIFAR, COCO and similar
datasets which include mainly images of objects and natural scenes
 So, extracted features mainly focus on color combinations, contours, shapes, etc.
3.3.4 August 2020 comparative study of different deep learning architectures for
classification of Chest X-rays: https://www.nature.com/articles/s41598-020-70479-z

You might also like