Do image recognition efforts always rely on machine learning and statistics? This is something I've always wondered.  Consider the Kinect.  It takes its 3d image data and manages to recognize that a human is contained at a given boundary.  Are these types of technologies exclusively machine-learning driven?
 A: No, or at least I would say not necessarily explicitly. If you have an image formation model (e.g. derived from the physics of the imaging process), you can pose recognition, reconstruction or detection as an inverse problem using parametric or implicit representations of your "pattern" or object of interest without making any probabilistic modeling explicit. 
For a more practical example, backprojection is a computationally efficient algorithm that solves the inverse radon transform and is often used to obtain tomographic pixel reconstructions (~ recognition of an image representing the object scanned). This is a situation where you have a well-posed inverse problem for a known forward model. 
That said, many inverse problems can be understood as bayesian MAP or ML inference problems, where the forward model is re-written as a probabilistic model.
For example, if the inverse problem is ill-posed, it is common to use regularization methods (e.g. TV or TR) to make the numerical treatment easier. However, many regularizers can be understood in a bayesian sense as priors acting on the parameters that the inverse problem aims to recover.
A: Amusingly, there are also some image recognition research efforts that do not rely on mechanistic identification - either by machine learning, statistics, or other automated methods. Instead, they contract the identification efforts out to human beings - who are fairly good at some forms of recognition, using a service like the Amazon Mechanical Turk.
Clearly, this approach won't work if you need real time image recognition, but it is an intriguing idea.
A: Yes and No. Nothing is ideal. Uncertainties come from everywhere. There is no exact mathematical model for all. Even though there is, it takes long time to figure out. The easiest way is to fit into probabilistic models. Machine learning is the learning (training the parameters ) using statistics of the given samples. 
Enes
