Make & Model Recognition

Deep Learning for Make & Model recognition

How can Deep Learning recognize the differences?

– Teil 2 – Deep Learning for Make & Model Recognition

To reach a fine-grained classification is the main challange for the Make & Model recognition. The reason is indisputable the subtle differences between the different classes. A great data record is necessary to learn the miniture differences. Please remember that under conditions such as little time, little computing power or little data, is not feasible. Deep Learning is to learn for experience and examples, which are based on algorithms.

For example Transfer Learning is a Machine Learning technique, which rededicates the learned classifiers for new tasks. For CNNs (Convolutional Neural Networks) a base network with a base data set is trained to generate significance and characteristics. At the end of Transfer Learning stands a classifier, which is applicable for the data set.

This means, that for us it is less work than teaching a complete new network. And it is also a quite powerful tool, to teach a big target network. At the same time it also minimizes data-overfitting. Of course you can always use a different technique.

What are neural networks or even convolutional neural networks (CNN)

Neural networks are the base for Deep Learning and consist of several layers. Best is to imagine a family of models, which are a bit like the human brain. At least that was where Deep Learning got its inspiration from. Their functions depend on a great number of inputs.

Concerning Make & Model : the bigger the data set of classified images of cars the better.

A neural net are examples of non-linear assumptions, where the model can learn to classify even more complexe relations.

This means that instead of learning a computer every single possibility of car models, we feed the Deep Learning program with tons of existing images.

A neural network consists of artificial neurons. They are arranged in layers. There are three kinds of layers:

  • Input Layer: searches for pixels in an image
  • Hidden Layers: can recognize certain characteristics, aspects, shapes, textures, ect.
  • Output Layer

Neural networks with more than two “Hidden Layers” can be defined as Deep Neural Networks. The greater number in layers the more complexe pattern can be recognized.

An example of two layers and you can imagen the connections between the neurons as a pattern, which is read.

Leave a Reply

Your email address will not be published. Required fields are marked *