CarND-Traffic-Sign-Classifier-Project: A Jupyter Notebook repository from vefak

Traffic Sİgn Classifier

Data Set Summary Exploration

1. Summary of Data

The size of training set is 139196
The size of the validation set is 17640
The size of test set is 50520
The shape of a traffic sign image is 32x32x1
The number of unique classes/labels in the data set is 43
Random images from dataset

The distribution of images for each data set shown in bars (Data augmentation included)

Train

Valid

Test

Design and Test a Model Architecture

Pipeline of preprocess

I decided to convert the images to grayscale and normalized them to increase accuracy. It is more easy to train normalized images. And most of image processing functions accepts only grayscaled images
I made some preprocess steps to increase size of dataset. They are flipping, rotating and sharping
After the data augmentation:
- X_train size = 139196
- X_valid size = 17640
- X_test size = 50520
- Number of Classes = 43

Model Architecture

The model that I used based on LeNet which is developed by Yann LeCun. I just added 1 conv layer, 2 fully connected layer and change the order of pooling layers. Also I used dropout method to prevent overfitting.

Layer	Description
Input	32x32x1 Grayscale image
Conv1 5x5x6	1x1 stride, valid padding, outputs: 28x28x6
RELU
Max pooling	2x2 stride, 2x2 kernel outputs: 14x14x6
Conv2 5x5x16	1x1 stride, valid padding, outputs: 10x10x16
RELU
Conv3 5x5x6	1x1 stride, valid padding, outputs: 8x8x32
RELU
Max pooling	2x2 stride, 2x2 kernel outputs: 4x4x32
Flatten	Output: 512
FC1	Input: 521 Output: 256
RELU
Dropout	Keep prob = 0.5
FC2	Input: 256 Output: 128
RELU
Dropout	Keep prob = 0.5
FC3	Input: 128 Output: 64
RELU
Dropout	Keep prob = 0.5
FC2	Input: 64 Output: 43

Training

The below parameters are similar to the ones used in LeNet.

EPOCHS = 25
BATCH_SIZE = 128
learning rate = 0.001
Dropout Probability = 0.5
The name of model = vefakNet
Normalizer = Softmax
Optimizer = Adam

Results

My final model results were:

training set accuracy of %99.6
validation set accuracy of %95.7
test set accuracy of %95.1

If an iterative approach was chosen:

What was the first architecture that was tried and why was it chosen?

I used LeNet firstly which is recommended by instructors
What were some problems with the initial architecture?

There is no dropout layers which can cause overfitting.
How was the architecture adjusted and why was it adjusted? Typical adjustments

I added one more conv and two more fully connected layer. So the number of learning parameters are greatly increased. Then I added droput layer with keep prob =0.5. Also I tried it with 0.25, 0.40 and 0.70 but 0.5 is the best one.
Which parameters were tuned? How were they adjusted and why?

I tuned dropout, number of images in dataset, learning rate, batch size and epoch. Increasing batch size, dataset and epochs effected in goodway. I tried different value of learning rate and 0.001 is gave best result
What are some of the important design choices and why were they chosen? For example, why might a convolution layer work well with this problem? How might a dropout layer help with creating a successful model?

The dataset consist of conv layers. The best and easiest option to obtain the information from images is conv layers. Droput is a great technique for overfitting. It kills random weights.

If a well known architecture was chosen:

What architecture was chosen?

GoogLeNet
Why did you believe it would be relevant to the traffic sign application?

It has large size of parameters but thanks to its architecture it doesnot affect computational cost. GoogLeNet also known for classiying images well.
How does the final model's accuracy on the training, validation and test set provide evidence that the model is working well?

We can say that the model works well because the results are consistent among themselves.

Test a Model on New Images

1. The five traffic sign images to be tested

Here are five German traffic signs that I found on the web:

2. Predictions

Here are the results of the prediction:

Image	Prediction
Speed Limit (30km/h)	Speed Limit (30km/h)
Bumpy Road	Bumpy Road
Ahead Only	Ahead Only
No Vehicles	No Vehicles
Go straight or left	Go straight or left

The model was able to correctly guess 5 of the 5 traffic signs, which gives an accuracy of 100%.

Output of performance analysis:

INFO:tensorflow:Restoring parameters from ./lenet
Image 1
Image Accuracy = 1.000

INFO:tensorflow:Restoring parameters from ./lenet
Image 2
Image Accuracy = 1.000

INFO:tensorflow:Restoring parameters from ./lenet
Image 3
Image Accuracy = 1.000

INFO:tensorflow:Restoring parameters from ./lenet
Image 4
Image Accuracy = 1.000

INFO:tensorflow:Restoring parameters from ./lenet
Image 5
Image Accuracy = 1.000