The differences between Deep Learning and the human brain
– Teil 3 – Make & Model Recognition
But Stop! It is not that easy as it sounds to create a big neural network and to call it Artificial Intelligents (AI). You should not forget one or two aspects:
- The artificial neurons are firing completely different than the human brain
- The human brain has 100 billion neurons and 100 trillion synapses and operates with about 20 watt (enough to run a weak light bulb) – in comparison a Neural Network has only 10 million neurons and 1 billion connections on 16.000 CPUs (about 3 million watt)
- The brain is limited to 5 types of input data from 5 sensory organs
- Children do not learn what a car is by showing them 100.000 images, which are classified “car” and “no car”, but thats how Deep Learning learns
- Most probably we do not learn by calculating the parial deviance of each neuron in connection to our initial concept (because by the way we do not know how exactly we learn)
Convolutional Neural Networks (CNN)
The advantage of CNNs is, that not only the importance of the features, but they can also learn the features itself. Additionally the CNNs have a highly modern accuray for the generic image classification.
The most important operation of Convolutional Neural Networks are the convolution layers. Let us imagine a 32x32x3 image, where we “convolve” (“fold”) a filter of 5x5x3 over it (the filter has to have always the same depth as the input image), then the result is an activation marker of 28x28x1.
The filter searches for a specific object on the image, this means it searches for a pattern in the whole image by using only one filter.
And now we want that our Convolution Layer searches for 6 different objects. Therefore the layer has to have six 5x5x3 filters. Each filter searches for a certain pattern on the image.
By the way the convolution (folding) is a linear operation. When we want the convolution to eventually stop, then we must add a non-linear layer at the end of the convolution layer. This normally is a Relu.
Another probably important point is, that it is not relevant for the filter where in the image the pattern is located. The pattern will be recognized nevertheless.
Where is Make and Model used?
Make and Model recognition of cars is nowadays in great demand for automatic vision based systems, e.g. for traffic supervision or surveillance cameras for law enforcement. Because the admission controls for parking lots and buildings or even restricted areas increased over the last couple of years also due to the raised safety awarness. For such surveillance systems a video streaming can help the security personell to identify vehicles and follow the vehicle by means of make, model or color.
For law enforcement angency it is required that a car, which is connected to a crime or a suspicious vehicle, can be observed over several million traffic recordings. And sometimes only thanks to the description of eye witnesses. But if you already use Make & Model one can save personell and the arrest of the offender leads quicker to success.
Additionally Make & Model Recognition (MMR) can also be used at border crossings and electronic tollgates. Or also for traffic control and surveillance, for statistics of traffic density, etc.
In short, the applications for Make and Model are:
- surveillance and evaluation
- safety monitoring
- collection of traffic data
- law enforcement
- toll gates