aiacademy: 深度學習 Computer Vision and Convolutional Neural Networks

cv introduction

• 之前　Udemy 認真的筆記：）

kernel/filter

kernel/filter:examples

• Shift, Blur, Sharpening, Gaussian Filter, … etc.

Convolutional Neural Network for Computer Vision Applicaions

說到介紹　cnn 還是要再一次好好的介紹　CS231

• Topics

  1. CV: object recongnition
     ML: multiple kernel learning

  2. CV: image segementation
     ML: graphical model

  3. CV: face detection
     ML: multi-task boosting

     

  4. CV: action recognition
     ML: low-rank reconstruction

  5. CV: multi-view people counting
     ML: transfer learning

  6. CV: image matching
     ML: energy minimization

     

  7. CV: fine-grained object recognition ML: CNNs with co-occurrence layer

  8. CV: patch descriptor learning ML: CNNs with adaptive learning rate

  9. CV: gesture recognition ML: DNNs with adaptive hidden layer

     

  10. CV: face age estimation ML: CNNs for hierarchical regression

  11. CV: image co-segmentation ML: Unsupervised CNNs

  ![Imgur](https://i.imgur.com/lgkNdGx.jpg)
  

Conventional approach vs. Deep Learning

Conventional approach to object recognition

• Training phase

  • image collection
  • feature extraction
  • classifier training
  • trained classifier

• Training phase

  • test image
  • feature extraction
  • trained classifier
  • prediction

Features are the keys

抓特徵！！！

• off-the-shelf visual features

  • SIFT
  • HoG
  • Constellation model
  • DPM

• Features are the keys to recent progress in classification
• Are handcrafted features optimal ?
• The optimal features for classification in general vary from task to task, even from category to category

Conventional approaches vs. Deep Learning

Deep Learning = Learning hierarchical representations

Neural Network

Neural nerworks and neurons

A sigle neuron

這邊可複習　cs231

• CS231: Lecture 4

• CS231: Lecture 5

What is deep neural networks (DNN)

CNN Intro

• imagenet

• Deep Neural Networks

• Ordinary Feedforward DNN with Image

• Characteristics of Image

• CNN Structure

Convolutional neural networks

• Reduce # of parameters

  • local connectivity

    

  • weight sharing

    

• Increate # of parameters

  • CNN with multiple input channels

    

  • CNN with multople output maps

    

Putting them together

Convolutional Neural Networks

• input image
• Convolution (Learned)
• Non-linearity
• Spatial pooling
• Normalization

• Feature maps

  • Convolution(Learned)

    

  • Non-Linearity

    

  • Spatial pooling

    

  • Normalization

    

AlexNet (代表性！！)

Object Recognition (2012 ~ 2014)