Scribd Logo
Upload

Welcome to Scribd!

  • 5 views

    Lecture 01

    Uploaded by

    lmh2505400617
    The document discusses classical image classification methods and convolutional neural networks. It provides an overview of shallow models versus deep models for image classification. It also describes the typical architecture of a convolutional neural network including convolutional layers, pooling layers, and fully connected layers.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Lecture 01

    Uploaded by

    lmh2505400617
    5 views53 pages
    The document discusses classical image classification methods and convolutional neural networks. It provides an overview of shallow models versus deep models for image classification. It also describes the typical architecture of a convolutional neural network including convolutional layers, pooling layers, and fully connected layers.

    Original Description:

    a lesson for deep learning and it

    Original Title

    Lecture-01

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document discusses classical image classification methods and convolutional neural networks. It provides an overview of shallow models versus deep models for image classification. It also describes the typical architecture of a convolutional neural network including convolutional layers, pooling layers, and fully connected layers.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    5 views53 pages

    Lecture 01

    Uploaded by

    lmh2505400617
    The document discusses classical image classification methods and convolutional neural networks. It provides an overview of shallow models versus deep models for image classification. It also describes the typical architecture of a convolutional neural network including convolutional layers, pooling layers, and fully connected layers.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 53

    Lecture 1

    Image Classification : From Shallow


    Models to Deep Models
    Today’s Agenda
    • Classical Image Classification Methods

    • Shallow Architecture vs Deep Architecture

    • Convolutional Neural Network


    Classical Image Classification Methods

    • Features are key to recent


    Illustration of SIFT Feature Calculation

    progress in recognition

    • Multitude of hand-designed
    features currently in use

    • Such as SIFT Features


    Classical Image Classification Methods

    • Widely-Used Classifier

    • Multi-layered Perceptrons Illustration of SVM

    • Support Vector Machine (SVM)

    • Random Forests
    Classical Image Classification Methods

    • Researchers pay much attention for the engineering of


    feature design

    • Carefully designed features are efficient for specific task,


    however they may be not work well for a new task

    • Is there a better way for feature designation, such as


    directly learning feature from data?
    Convolutional Neural Networks
    • Convolutional Neural Network (CNN) is an efficient deep
    learning based method widely-used on lots of image-
    related tasks including image classification.

    • CNN is a special kind of neural network.

    • CNN apply convolution on input images and intermediate


    results to produce powerful feature representation.
    Now Days : CNNs are Everywhere
    Typical CNN Architecture
    • A typical CNN is comprised by :
    • Convolutional Layer with rectified linear units (ReLU)
    • Pooling Layer
    • Fully Connected (FC) Layer
    Rectified Linear Units (ReLU)
    • ReLU avoids or rectifies vanishing gradient problem

    • Only used within hidden layers of deep neural networks


    1 x 1 convolution layers are usually adopted for channel compression
    Max Pooling & Average Pooling

    6 8
    Max Pooling
    1 1 2 4 3 4

    5 6 7 8 Pooling with
    2x2 filters
    3 2 1 0 and stride 2

    3.25 5.25
    1 2 3 4 Average Pooling
    2 2
    Fully Connected (FC) Layer
    • FC Layer fully connects all inputs with outputs, as in ordinary neural
    networks

    • It is usually adopted at the last layer with a softmax activation


    function

    You might also like