1 of 14

Greendeck Image Classification

Capstone Project: Group 11

Mukul Agrawal

Vinay Gupta

Sheldon Mascarenhas

Dr. Anjali Rauut

Dwipayan Pathak

2 of 14

Problem Statement

GreenDeck, help retailers with pricing intelligence. One of our main challenges comprise of classifying the products into right categories and matching the exact same products across different retailers for price comparison.
This problem comes under the category of multi-label classification and product retrieval. Sometimes product's textual data (product title & description) is helpful for this task, and sometimes product's images help. Hence, we need a Deep learning model for classification tasks and product retrieval.
Group needs to build an Image Classifier using ML techniques on Python to enable the model to identify images

3 of 14

-:APPROACH:-

Know your Train and Test data
Data Split
Preliminary EDA

EDA

VGG16 Model
Data Compress and Feature Extraction

FEATURE EXTRACTION

Building Sequential Model for Training
Model fitting and accuracy check

TRAIN MODEL

Model Accuracy on Test Data
Conclusion

TEST MODEL

4 of 14

Know your Train and Test Data: EDA

Command (len) tell us that there are 6000 images store for analysis

With pd.unique command we can see that below are the unique categories

array(['shoes', 'beauty', 'jewelry and watches', 'bags'], dtype=object)

We assigned integer code to each of these categories so that it can be converted into hot encoded vector

5 of 14

Data Split: EDA

Take away- We have split the data for each class into 70-30 with train_test_split method so that we can have to 2 separate data base for training and testing purpose

6 of 14

Preliminary analysis: EDA

Take Away – Above graph shows that sample available for images are sufficient to train deep learning algorithms for image classification

7 of 14

Preliminary analysis: EDA

Take Away – To run CNN-VGG, image needs to be in compressed form (224*224 pixels), which is a pre-requisite to complete the classification of Images

8 of 14

VGG16 Model: FEATURE EXTRACTION

Take Away – VGG 16 model is used to extract out the features from the images in the form of Matrix.

9 of 14

Data Compress and Feature Extraction:� FEATURE EXTRACTION

To work with CNN-VGG, image needs to be in compressed to 224*224 pixels.
Features extraction have been done using VGG16.

10 of 14

Building Up A Sequential Model: TRAIN MODEL

Take Away – Using model transfer. we will try to build a sequential model

11 of 14

Model fitting and accuracy check: TRAIN MODEL

Take Away – Using EPOCH =20 we will check the accuracy of the model.

Here you can see we reached 99% accuracy

12 of 14

Model Accuracy on Test Data: TEST MODEL

Take Away – When we put the model in use with Test data available (30%), we can see that the accuracy is at 96%.

13 of 14

-:Conclusion:-

We used VGG 16 to conclude our model accuracy.
With 96% accuracy we can safely put the model in production to categorize items related to fashion.
This will help in categorizing the images in right categories for different retailers to do price comparison.

14 of 14

Thank you!! ☺

This Photo by Unknown Author is licensed under CC BY-SA-NC