Original article was published on Deep Learning on Medium
Traffic Sign Classification using Residual Networks(ResNet)
Deep Residual Learning to Classify Traffic Signs
Deep Convolutional Neural Networks(CNNs) are widely used to solve various computer vision tasks in the field of Artificial Intelligence. This article focuses on developing a deep learning model in order to recognize traffic signs.🛑❌🚫🚷🚳
Table of Contents
- Data Analysis
- Create a ResNet Model
- Model Training
- Model Evaluation
First of all, we need a dataset to train the deep learning model to recognize traffic signs. Kaggle Datasets is the best platform to find datasets for different tasks. Such as Machine Learning(ML), Deep Learning(DL), and Data Science.
Here is one of the datasets contains nearly 73,139 diverse images of traffic signs of 43 classes.
In this section, we are going to use a simple way to analyze the dataset.
Here is a simple count plot to analyze the spread of data in classes. The below code is used to plot the graph:
Let’s visualize some of the samples from the dataset. This will help us to understand the data. The below code serves the purpose by plotting 100 images from the dataset.
Create a ResNet Model
In this section, we are going to create a deep learning model to recognize traffic signs.
Microsoft introduced the deep residual learning framework to overcome the ‘degradation’ problem which is a hard optimization task. The shortcut connections i.e., skipping one or more layers as shown in the below figure.
These shortcut connections perform identity mapping and the outputs are added to the outputs of stacked layers. This has solved many problems such as :
- Easy to optimize
- It gains accuracy from greatly increased depth, producing results that are better than previous network architectures.
For a better understanding of deep residual learning. Use the research paper entitled ‘Deep Residual Learning for Image Recognition’ which is freely available on arxiv.
We are going to use the TensorFlow applications module which provides different popular deep learning models with pretrained weights to use.
We are going to use ResNet50 architecture without pretrained weights. We add the dense layer with softmax activation at the end to predict the classes. Below is used to create the model.
You can see the visualization of the model created using the plot_model method.
These are the parameters used during the training process. The batch size as 32, epochs 50, learning rate as 0.001, loss metric ‘Categorical Cross Entropy’, optimizer as ‘Adam’. The callbacks ModelCheckpoint, EarlyStopping, ReduceLROnPlateau, and CSVLogger are used in the training of the ResNet50 model. You can use the below link for understanding the nuts and bolts of callbacks.
The below code is used to compile and fit the model.
The graphs between accuracy over epochs on training and validation data.
The graph between loss over epochs on training and validation data.
You can see the loss and accuracy converges after 20 epochs.
Let’s see the classification report which helps to evaluate the model.
The output results in the form precision, recall, F1 score with respect to each class.
The confusion matrix is used to describe the performance of the classification model. The below code is used to generate confusion matrix:
The resultant confusion matrix is shown below:
The classwise accuracy can be derived using the below code:
Few samples from the unseen data are used for predicting the class labels using the trained ResNet50 model. The below code is used for this purpose:
The prediction of the unseen data is shown below:
The code that I have written for the task is available as Kaggle Notebook. Feel free to use it. Here is the link: