TASK 3 --> Computer Vission

Dataset Acquisition

Took about 205 screenshots of people in mask.

Approach to Create dataset

• Searched for images of people with mask on on google.Found a few sites where I could find high quality straight headshot photograph of a person wearing a face mask.
• I tried to include as many different ethnicities, races, colors that I could find so that the model is able to classify the general population
• Ensured to make a balanced dataset with equal number of images of females and males,so that my dataset is not skewed towards one group.
• Also deliberately added images of ladies in sarie, wearing bindi and men wearing gamcha so it is better able to classify indian population.

Approach to the Problem

• Using custom model:
  • In the begging I used a simple CNN model , that consisting of 3 alternating Conv2D and Maxpool layers pair, followed by flatten and dense layers, the model was overfitting so I used Data augmentation, to flip the image randomly along the vertical axis, rotate the image and doing a random zoom , also I added a dropout layer.
  • As a result, I was able to prevent overfitting and was at best case after several epochs able to reach accuracy of 86%.
• Using Transfer learning:
  • I used the InceptionV3 model(Excluding its last 2 layers) trained on ImageNet dataset as the base of my model, with its training stopped. On top of that I added 3 Dense layers of my own which would serve the purpose of classifying particular to my dataset.
  • Why I used Transfered learning?
    • The InceptionV3 is already trained on millions of images, so its CNN network is better able to detect any edges,contours, shapes than my custom trained model.
    • So I am using this CNN network to analyse the image better,and then feeding the dense layers with this more accurate description, which is as a result increasing my accuracy.
  • I was able to achieve ~92% with Transfer learning.

Roadblocks and Measures to Increase Accuracy

During the task, I encountered several challenges:

• Finding images:
  • It took extra effort to get high quality images of people of different ethnicity , color, etc as I had to make a dataset that would represent different kinds of people.
  • Many images had watermarks and sites had paywall, so had to overcome those.

To increase accuracy, I:

• Fine-tuned model hyperparameters using cross-validation.This includes changing how much I want to augment my data,what should be the patience of early stopping ,selecting the batch size ,etc.
• Used tranfer learning to use InceptionV3 model as the base of my model.

Instructions to Run the Code

Link to Hugging face app --> Link
Link to Weights(too large to be uploaded to GitHub) --> Link

To run the code:

1. Clone this repository to your local machine.
2. Upload the task_3___Aarush_singh_kushwaha copy.ipynb in colab and read the comments as a guide to run the notebook.
3. Upload the Female.zip and Male.zip dataset to colab.
   
4. Upload the model_weights.h5 to enjoy the use of the saved weights.
   NOTE: These weights belong to the Model made with InceptionV3 (ie. the second model)
   OR

Directly run in colab though this Notebook link

Google Drive links:

• Male Images Dataset
• Female Images Dataset

Screenshots