Digit Classification using Logistic Regression and CNN

This is a project focusing on the classification of digits from 0-9 using Logistic Regression and Convolutional Neural Network. The trained model is used to predict digits drawn on the captured frames from the webcam using object tracking.

Getting Started

Built with

The Logistic Regression is implemented using numpy and softmax function is used for multiclass classification. The Convolutional Neural Network uses Keras API with tensorflow as backend.

Pretrained weights

Training examples = 8000, Test examples = 1000

Logistic Regression

Train accuracy = 92.1%, Test accuracy = 91.3%, learning rate = 0.1

CNN Model

Train accuracy = 98.1%, Test accuracy = 96.2%, no. of epochs = 8

(Feel free to train the models on your own.)

Prerequisites

Install Conda to resolve all requirements of python related dependencies.

Usage

Files usage

• LRmodel.py : Logistic Regression implemented using numpy
• conv_network.py : CNN model implemented using Keras API
• digit_recognizer.py : File to either train or load saved weights
• camera_pred.py : Used to test the models using webcam

Training the models

To Train the models on your own, delete the weights folder and run digit_recognizer.py

python digit_recognizer.py

(If any of the files in weights folder is not present, the models will be trained again.)

Testing using Camera

Run camera_pred.py (Use a green colored object to draw digit inside the red box).

python camera_pred.py

Press c to clear the box.