/Understanding-CNN-and-Neural-Style-Transfer

Visualizing CNN and Neural aesthetics

Primary LanguageJupyter Notebook

Visualizing CNN and Neural aesthetics

Slide about "Visualizing CNNs and Neural aesthetic" topic.

This repository contains Understanding CNN by visualizing & Neural Style Transfer tutorial notebook. I further provide an overview of the current progress of NST topic, which is receiving increasing attention from both in academic literature and industrial companies.

The src/ folder contains experiment TensorFlow code for Visualizing CNN (Using Inception_V1 trained model), Fooling CNN by adversarial examples, Refining Neural image & Deep dream.

Some results from these experiments:

Neural Style transfer

Oak Tree styled by dwelling image

Ha Long Bay styled by eva image

Fooling CNN by Adversarial noise

Predicted as Kelpie
Confidence: 0.872 		Adversarial noise Predicted as Giant panda
Confidence: 0.996

Refining neural image

Tensor Flowers




Deepdream


Reference materials:
1. Stanford CS231n - Visualizing and understanding CNN
2. Yongcheng Jing et al.(2018) - Neural Style Transfer: A Review
3. Christian Richardt, Rainbow Group - Non-Photorealistic Rendering (NPR)
4. Deeplearning.ai Course 4
5. Udemy Advanced Computer Vision
6. Inceptionism: Going Deeper into Neural Networks