https://github.com/j-bodek/formal-photo-verification

An easy web application that predicts whether the photo is in a formal style
https://github.com/j-bodek/formal-photo-verification

computer-vision convolutional-neural-networks fastapi resnet-50 streamlit-webapp tensorflow2

Last synced: about 2 months ago
JSON representation

An easy web application that predicts whether the photo is in a formal style

• Host: GitHub
• URL: https://github.com/j-bodek/formal-photo-verification
• Owner: j-bodek
• Created: 2021-08-30T19:23:51.000Z (about 4 years ago)
• Default Branch: master
• Last Pushed: 2021-09-01T10:48:29.000Z (about 4 years ago)
• Last Synced: 2025-06-06T03:41:48.461Z (5 months ago)
• Topics: computer-vision, convolutional-neural-networks, fastapi, resnet-50, streamlit-webapp, tensorflow2
• Language: Jupyter Notebook
• Homepage: https://share.streamlit.io/liljack118/formal-photo-verification/app/app.py
• Size: 87.9 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Formal Photo Verification App 

This easy web application predicts whether the photo is in a formal style. 


APP: https://share.streamlit.io/liljack118/formal-photo-verification/app/app.py 


API: https://formal-photo-verification-api.herokuapp.com/

## Data 

To train network I used dataset of 3500 images in total. 


Training set - 2800 images ( 1200 positive / 1600 negative) 


Validation set - 350 images (150 positive / 200 negative)


Test set - 350 images (150 positive / 200 negative) 


Invalid images contains:

- rotated head 

- frontal head photos with sunglasses/hats/headset/unnatural mimic

- frontal head photos on wrong background

- animated frontal head photos

- empty rooms photos

- person with body 

 

Invalid images from every subfolder were distributed equally to training,validation and test set to make sure that images in every set came from same distribution.

![](model/images/1.jpg)

![](model/images/2.jpg)

## Model

To make predictions I used ResNet50 network with imagenet weights. Firstly i extracted features (that will be feeded to fully connected layers) of headless model and then visualize them with t-sne to see if they are well seperated. Input size is (224,224,3).

![](model/images/3.jpg)

![](model/images/4.png)

As we can see ResNet50 with imagenet weights makes good job to separate valid and invalid images. Now let's add fully connected layers to make predictions! I chose to add two instead of one because it performs slightly better. 

![](model/images/5.jpg)

Model performance on test set: 


![](model/images/6.jpg)

## APP

To create app i used streamlit library and deployed it on https://share.streamlit.io/. 

## API

Api was created with fastapi and deployed on heroku.