CNN does support an extensive amount of use cases, not only computer vision one. Nevertheless, it underperforms with variations e.g. rotations.
The convolutional layers in front of the fully connected layer act as filter, as dot-product with a field e.g., 3×3 of features. The filter overcome the usual sparsity of natural data. Sparsity describes the “not existence” of information e.g., with LIDAR data one some points have a value (> or < 0).
But stop, often not take under consideration are the pooling layer (combining features), but these layer are highly important. They perform the downsampling e.g., dimensional reduction from 3D (height, width, color) to 2D (height, width, black/white) or reduce the number of feature from 800×600 to 400×300.
One reason why CNN are so powerful is the fact, that the usual feature engineering is “not” required. The filter and pooling identifies orders e.g., vertical structures.
A nice evolution overview:
And a bit of taxonomy
