Gesture-Recognition-CNN

Cotents

About
The Project
The Process
Data Collection
Data Preprocessing
CNN Model
Implementation
Results

About

This Project was made to understand the concept of CNN and to learn about the various layers of CNN. This project is purely done in numpy (no modules even for backpropogation)

The Project

The project basically recognizes gestures from one of the below gestures

• FIST
• HAND
• ONE
• PEACE

The Process

• First the data is generated from scratch using image processing
• The data is cleaned and a split is made, later the model is pickled
• CNN model is built using only numpy
• Model with different parameters are trained on google cloud
• Models are compared

Data Collection

OpenCv is used for image processing, live video was captured via video camera, then with each frame

• Image is converted in gray scale
• Gaussina blur is applied
• Mask is applied
• Image is dilated and eroded
• Median blur is applied and the image is saved

The final image that was saved was a 50x50 gray scaled image

The images can be found in the images_data folder

Images from the dataset

Fist

Hand

One

Peace

Here is the link for the data generation code

Data Preprocessing

This step preprocesses the data i.e normalizing and splitting the data into train, test, validation. The split ratio is 70%,20%,10% repectively. The data is later pickled so as to load the object directly
The preprocessing code can be seen here

Total data points (approx)
train - 3000
validation - 500
test - 1000

CNN Model

This section describes the architechture used, the next section explains the parameters in detail and how the model was trained on the cloud
The model consisted of two CONV BOX followed by a fully connected layer with two hidden layer

One CONV BOX consists of three layers
Convolution Operation -> Relu -> Max Pooling
So, this operation was carried out twice

Convolution operation was calculated using the numpy's fast fourier transform
Max Pooling was implmented by using loops and some fancy array slicing

After Conv box, Fully connected layer was attached - fully connected layer consists of two hidden layers

The Entry point to the class is the train function

The CNN Model class can be viewed here

Links used to learn CNN

• https://www.youtube.com/watch?v=FTr3n7uBIuE
• http://cs231n.github.io/convolutional-networks/
• https://ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/
• http://colah.github.io/posts/2014-07-Conv-Nets-Modular/
• http://andrew.gibiansky.com/blog/machine-learning/convolutional-neural-networks/

Backpropogation

• http://cs231n.github.io/optimization-2/
• https://grzegorzgwardys.wordpress.com/2016/04/22/8/
• https://www.slideshare.net/kuwajima/cnnbp
• http://www.jefkine.com/general/2016/09/05/backpropagation-in-convolutional-neural-networks/

Implementation

This section provides some Implementation notes.

• Implemented Dropout to avoid overfitting
• Applied gradient check to verify the gradient update
• Used scipy optimize class to minimize the loss function
• Xavier initialization for initial guess of parameters
• Trained the model parallely accross different cores (using joblib)

Trained the model on Google cloud platform

Google Compute Engine Configurations

• Ubuntu 16.04
• 8 CPUs
• 50Gb ram

Once the model was trained it was pickled and this code is used to predict the gesture

Results

All the parameters related to the model trained was written in params file, the value of the parameters was read during the run time
Around 15 models were trained, I will be comparing here 4 models, based on the major changes

No.	# Kernels L1	Dimension L1	# Kernels L2	Dimension L2	# Hidden Nodes L1	# Hidden Nodes L2	Optimization Method	Validation Accuracy %
1	3	2x2	4	4x4	300	150	TNC	16.04
2	16	9x9	32	7x7	800	400	L-BFGS	22.32
3	32	9x9	64	5x5	3000	1500	BFGS	60.67
4	32	3x3	64	5x5	1800	900	TNC	95.71

Note - training time for every model is 4+ hours

The 4th model in the table gave the best results, currently params.json contains the parameters of the best model
The model was trained for 100 iterations

Below is the figure showing the loss vs iterations curve for the best model

Finally the model was tested on the test set and the accuracy was 95.06%

Also the model was tested on live feed of the web cam it was giving impressive results, one peculiar observation was that the model was getting confused between hand and the fist. Also the model can perform better if the model is trained for more number of iteration (the current model is trained for 100 iterations)