Traffic Sign Detection Sweden

My solution to Kaggle Project to detect Traffic signs in Sweden and Classify them

Dataset

The dataset is made up of 3113 traffic sign images belonging to Sweden. The Kaggle competition from which the dataset was picked, divided the dataset into a training set of 2503 images and a testing set of 610 images. Each image consists of different dimensions in terms of height and width but all the images are made up of 3 color channels, i.e, RGB (Red, Green, Blue). Further, the intensity of light and pixel resolution of each image varies independent of the class which sometimes leads to the images being too difficult for a human to recognize. The complete dataset consists of 17 classes in total and each class represents a traffic sign in common except for one class named ”others” which represents traffic signs that does not belong to the other 16 unique classes.

Data Augmentation

The dataset provided for training will not be sufficient for a model to generalise well. Figure 1 represents the class distribution in the training dataset. Most of the classes have less images for a model to train on.

Therefore, the dataset is expanded by flipping the images so that the flipped images still represent the same class after flipping.

Similarly, images belonging to class ”30 SIGN” are invariant to vertical flipping. Hence, flipping it vertically resulted in an image as shown in figure 3. Therefore, images that are invariant to horizontal or vertical flipping were transformed to extend the dataset size while keeping the number of classes constant.

Finally, the initial training dataset of 2503 images is expanded to a final training dataset of 4917 images.

Network

To solve the problem at hand, we chose a Residual Network variant which consists of 50 layers, i.e, ResNet-50. The motivation behind considering such a network is due to the fact that ResNet-50 is fairly faster to train yet so powerful. Choosing a more deeper network could result in a higher accuracy but needs a significant amount of time to train along with a powerful Graphical Processing Unit (GPU).

Results

ResNet-50 model was successfully trained to classify Sweden traffic signs. The model trained on augmented dataset has achieved an F1-score of 84%.

Precision	Recall	F1-Score
0.88	0.80	0.84

Although the dataset was very small, we could effectively train the powerful ResNet model to achieve the desired results with the help of transfer learning. There is more scope for performance improvement by further increasing the dataset size using various data augmentation and also by implementing different preprocessing techniques.