Anthropomorphism.pptx

Anthropomorphic Animal Face Masking using Deep Convolutional Neural Network by

Rafiul Hasan Khan

3/28/2019

MCSP

1

 Contents • Introduction • Dataset • Proposed structure • Simulation

• Result • Layer inspection • Proposed method for morphing

3/28/2019

MCSP

2

 Introduction • Classification of ‘Bear’ ,‘Cat’ & ‘Deer’. • Large Dataset of quality images. • 12 layer Deep Convolutional Neural Network. • Achieved high accuracy rate with a low computing cost. • This network can be applied to any class for binary classification.

3/28/2019

MCSP

3

 Dataset • Dataset is made of two classes 1. Bear 2. Cat 3. Deer

• Each class contains one thousand images. • Images were collected and cropped from various sources. • All the images were resized by 227x227

• Images were labeled by the names of their classes.

3/28/2019

MCSP

4

 Dataset • Example of Dataset

3/28/2019

MCSP

5

 Proposed Structure

Fig : Structure of the network 3/28/2019

MCSP

6

 Schematic Diagram

3/28/2019

MCSP

7

 Simulation • The Dataset was divided into two parts of eighty by twenty percent for ‘Training’ and ‘Testing/Validation’. • Initial Learning rate was set to 0.001 • Number of ‘Epoch’ was set to 20. • Mini Batch Size was set to 64.

3/28/2019

MCSP

8

 Result

Fig : Classification result of the Dataset 3/28/2019

MCSP

9

 Layer Inspection • Input Layer

Fig : Unscaled Input Image

3/28/2019

Fig : Scaled Input Image by 227x227

MCSP

Fig : Data Layer activations

10

 Layer Inspection • First Convolution Layer

Fig : Convolution Process

Fig : Activation of first Convolution Layer (128) 3/28/2019

MCSP

11

 Layer Inspection • First Convolution Layer

Fig : Strongest activation Channel of First Convolution Layer 3/28/2019

MCSP

12

 Layer Inspection • First ReLU Layer

Fig : Activation of first ReLU Layer (128) 3/28/2019

MCSP

13

Layer Inspection • First ReLU Layer

Fig : Strongest activation Channel of First ReLU Layer 3/28/2019

MCSP

14

 Layer Inspection • First Pooling Layer

Fig : Activation of first Pooling Layer (128) 3/28/2019

MCSP

15

 Layer Inspection • First Pooling Layer

Fig : Strongest activation Channel of First Pooling Layer 3/28/2019

MCSP

16

 Layer Inspection • Cross Channel Normalization Layer

Fig : Activation of Cross Channel Normalization Layer (128) 3/28/2019

MCSP

17

 Layer Inspection • Cross Channel Normalization Layer

Fig : Strongest activation Channel of Cross Channel Normalization Layer 3/28/2019

MCSP

18

 Layer Inspection • Second Convolution Layer

Fig : Convolution Process

Fig : Activation of Second Convolution Layer (384) 3/28/2019

MCSP

19

 Layer Inspection • Second Convolution Layer

Fig : Strongest activation Channel of Second Convolution Layer 3/28/2019

MCSP

20

 Layer Inspection • Second ReLU Layer

Fig : Activation of Second ReLU Layer (384) 3/28/2019

MCSP

21

 Layer Inspection • Second ReLU Layer

Fig : Strongest activation Channel of Second ReLU Layer 3/28/2019

MCSP

22

 Layer Inspection • Second Pooling Layer

Fig : Activation of Second Pooling Layer (384) 3/28/2019

MCSP

23

 Layer Inspection • Second Pooling Layer

Fig : Strongest activation Channel of Second Pooling Layer 3/28/2019

MCSP

24

 Layer Inspection • Output Layer

Fig : Activation of Output Layer 3/28/2019

MCSP

25

 Layer Inspection

Fig : Unscaled Input Image

Classification

Bear

Cat

Deer

………..

Prediction

4.47e-09

1.0

7.02e-11

…………

Fig : Result of Input Image 3/28/2019

MCSP

26

 Proposed Method for Morphing • First Convolution Layer

 For the morphing process, the control points will collected from the Strongest activation Channel of First Convolution Layer. The channel will go through preprocess to be eligible for Control Point Selection process. Fig : Strongest activation Channel of First Convolution Layer 3/28/2019

MCSP

27

3/28/2019

MCSP

28