Lab 5

Link to source code

Team Members:

Name: Avni Mehta, Class Id: 11 Name: (Trinadha) Raji Muppala, Class Id: 14

Objective:

Compare the accuracy of the Image classification using CNN and Inception model for VQA interiors data (bathroom, bedroom, living, kitchen). Using FLASK build website for object detection.

Input:

MS Coco data for interiors (bathroom, bedroom, living, kitchen). Training data for 50 images and testing data 5 images for each category In order to perform CNN on our project dataset, we had to perform the following pre-processing steps:

Data Pre-processing:

Explanation: In order to perform the CNN on our project dataset, we had to perform the following pre-processing steps: • Resize all the training and test images to 28x28 pixels.

• Change the image colour from RGB to black and white.

• Convert the resized images to MNIST format. Reference

Output is:

For Inception used the images as it is.

For FLASK used the model generated by Incdeption

Code:

CNN model with 2 hidden layers active map size 16, 36 output layer 128

Inception model Code:

FLASK code

Index.html

Output:

Accuracy for image classification comparison with CNN Adamoptimizer (500 and 100 epochs) ,Adagradoptimiser with 500 epochs and Inception model

Adamoptimizer accuracy 100 epochs

Adamoptimizer accuracy 500 epochs

Adagradoptimiser accuracy 500 epochs

TensorBoard for CNN

Inception model accuracy output

Inception model tensorboard

Website using FLASK object detection output

Correct prediction (out of 20 test images 5 for each category except 2 images all got predicted correct )

Wrong Prediction (out of 20 test images 5 for each category these two predicted wrong)

FLASK python terminal output