guest talk- roof classification

Post on 15-Apr-2017

58 Views

Category:

Data & Analytics

6 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Deep Learning : Roof Style Classification

problem

Kishore Kumar MohanM.S in Information Systems Student

Under the guidance ofProf. Sri Krishnamurthy

Roof style Classification using convolutional neural network

2

Objective: To classify various roof styles using satellite aerial images captured from Google Earth and to benchmark accuracy obtained using regular neural nets and convolutional nets.

Problem Type: Supervised Classification

Quick Overview of Data:

Roof type 01- Flat: Roof type 02- Gable:

Roof type 03- Hip: Roof type 04- Gambrel:

Business Impact:

- Underwriting heavily relies on roof type for certain states. Third party data

vendors like Black knight http://www.bkfs.com/data-and-

analytics/Pages/default.aspx

- Infact some insurance companies to make data collection easier provides

discount on premium for clients if they choose roofing type of their choice.

https://www.statefarm.com/insurance/home-and-

property/homeowners/discounts/roofing-materials

Packages Used

4

Python Library – Tensorflow (Bitfusion – Amazon AWS), Opencv

Why Opencv and not scipy - Improved Gaussian blur performance.

Approaches and Reading Images

5

- Regular Net

- Regular Net with edges detected

- Convolutional Neural Net

Reading Images for all 3 approaches:

Regular Net (1x16384)

Regular Net – Edges(1x16384)

CNN(128x128x3)

Regular Net – Code and Math

6

Condition for matrix multiplication (X.W): we should have "mxn" and

"nxj" format to get "mxj" form. We break the image to 1x16384

Convolutional Neural Net – Code and Layers

7Optimizer – SGD with learning rate = 0.001

Activation functions - CNNRelu Logit

- Imitates Biological neural nets. Activation occurs only when the input signal strength is greater than threshold

- Fine values between 0 and 1. Less computational complexity than softmax

8

Accuracy and Analysis

9

Summary

• Reading images for a specific architecture

• Why opencv?

• Regular Net – Math and code

• Feeding images after extracting features has zero impact

• Convolutional neural network and activation functions

• Accuracy and Analysis

10

top related