traffic_sign_classifier

This project contains the tensorflow based neural network for classification of Traffic signs in German Traffic Sign dataset

Dependencies

• Python 3.x
• Matplotlib
• Sklearn
• Tensorflow 1.x

Steps:

Loading the Dataset:

The pickle files of dataset were hosted on the google drive and then drive was mounted on the colab notebook.

Exploring the Dataset:

With the help of Python and Numpy commands the dataset was explored. The summary of exploration is as follows

• Number of images in training set is 34799
• Number of images in validation set is 4410
• Number of images in test set is 12630
• The shape of a traffic sign image is (32, 32, 3)
• The number of unique classes/labels in the data set is 43

Data Preprocessing:

Currently only normalization and gray scale conversion techniques are being applied.

Network architecture design:

A slightly modified version of LeNet was used, details of the network and training can be found below.

Layer	Description
Input	32x32x1 grayscaled, normalized image
Convolution 5x5	1x1 stride, valid padding, outputs 28x28x6
RELU
Max pooling	2x2 stride, outputs 14x14x6
Convolution 5x5	1x1 stride, valid padding, outputs 10x10x16
RELU
Max pooling	2x2 stride, outputs 5x5x16
Flatten	5x5x16 -> 400
Dropout	keep_prob = 0.8
Fully connected	400 -> 120
RELU
Dropout	keep_prob = 0.75
Fully connected	120 -> 84
RELU
Fully connected	84 -> 43

To train the model, I used Adam optimizer. The final settings utilized for training are given below:

Parameter	Value
Batch size	128
Epochs	50
Learning rate	0.001
Mean, mu	0
Variance, sigma	0.1
Dropout keep probability	0.80 (layer 1), 0.70 (layer 2)

Results:

My final results include,

• Training accuracy of 98%
• Validation accuracy of 94%

Initial results showed lot of overfitting as validation accuracy was stagnant @ 87%, fine tuning of dropout layers boosted the validation accuracy. Below are the graphs of Training/Validation accuracy/losses vs Epochs.

Improvements:

• Overfitting can still be reduced
• Synthetic data generation can be performed to balance the samples per class.