[tokipin]

Take this[1] for example. That's a regular Mathematica notebook. You can insert images (like drag & drop them from your desktop) directly into notebooks. It's actually very convenient for image processing. The same thing holds true for things like the graphical output of plotting functions, which retain their underlying vector data.

The symbolic paradigm isn't necessary for this, but it makes it easier. Implementation-wise these things requires less "special case-ing" than you would think because the symbolic paradigm in essence encourages your representations to be naked. Thus images, plots, etc are not "black box" items. They're structured data represented in the same language, but which the frontend happens to display in visual forms.

(Code itself is also data, except that if a symbol in the data has an associated transformation rule, that transformation rule is applied. This mechanism is how you make data do 'variables' and 'functions'. These variables and functions work more or less the same as they would in a common programming language, but the underlying mechanism is nothing but data transformations.)

[1] http://reference.wolfram.com/mathematica/ref/EdgeDetect.html