Motivation for Convolutional Neural Networks
Finding good internal representations of images objects and features has been the main goal since the beginning of computer vision. Therefore many tools have been invented to deal with images. Many of these are based on a mathematical operation, called convolution. Even when Neural Networks are used to process images, convolution remains the core operation.
Convolutional Neural Networks finally take the advantages of Neural Networks (link to Neural Networks) in general and goes even further to deal with two-dimensional data. Thus, the training parameters are elements of two-dimensional filters. As a result of applying a filter to an image a feature map is created which contains information about how well the patch corresponds to the related position in the image.
Additionally, convolution connects perceptrons locally. Because features always belong to their spatial position in the image, there is no need to fully connect each stage with each other. Convolving preserves information about the surrounding perceptrons and processes them according to their corresponding weights. In each stage, the data is additionally processed by a non-linearity and a rectification. In the end, pooling subsamples each layer.
Deep learning finally leads to multiple trainable stages, so that the internal representation is structured hierarchically. Especially for images, it turned out that such a representation is very powerful. Low-level stages are used to detected primary edges. High-level stages lastly connect information on where and how objects are positioned regarding the scene.
Figure 1: Typical convolutional neural network with two feature stages .
After introducing relevant basics in image processing and discrete convolution, the typical layers of convolutional neural networks are regarded more precisly.
A typical convolutional neural network is composed of multiple stages. Each of them takes a volume of feature maps as an input and provides a new feature map, henceforth called activation volume. The stages are consecutive separated in three layers: A convolutional layer, a ReLU layer and a pooling layer. The fully-connected layer finally maps the last activation volume onto a class of probability distributions at the output.
The following chapters will provide an overview regarding the structure and the tasks of each layer.
 Y. Ma, S. Soatt, J. Kosecka, S.S. Sastry. An Invitation to 3-D Vision: From Images to Geometric Models. Springer New York, 2005.
 Y. LeCun, K. Kavukvuoglu, and C. Farabet. Convolutional networks and applications in vision. In Circuits and Systems, International Symposium on, pages 253–256, 2010.
 Y. LeCun, Y. Bengio, G. Hilton. Deep Learning. Nature 251, pages 436-444, May 2015.