With very low harmonics you can see that it is not possible to encode a very fast varying shape.. the only way for having resolution is to use high harmonics..
Consider a basic idea of the High Frequency of an event X. If X has high frequency it means it changes very fast or in other words it has rapid transitions over a given time interval. Now assume that X is a location where there are edges in an image. You should first understand the fundamental principle at the edges of an image. There are abrupt rapid changes of the intensity values at the edges of an image. You can check this by observing the intensity values of an image at both flat and edge regions (you may use MATLAB to read a very simple image and display the intensity values). In the context of frequency, therefore, the regions X where there are rapid transitions intensity values or abrupt color changes have high frequencies. The term "spatial frequncy" just means that we are in the real world time axis or the image is in its natural state (not transformed). For audio signals we can easily tell if the signal has high frequency by listening, but for images the phenomenon can be explained by seeing.
You can describe best with Fourier transforms in a mathematical sense but it means in a basic way, the edges are changing sharply in small space, so high frequency, just as very small things require high spatial frequencies to describe them in frequency space.
One of my undergraduate students group did similar workaround 9 years back: optical character recognition to recognize car chassis number (which is embossed on metal and hence font is same colour as the background) using ANN. We even published a paper on that work a little later. Please refer to it and share your queries.
Parul Shah, Sunil K., Taskeen N., G. Nikita, K. Kaushik, L. Ketan, “OCR-based Chassis-
Number Recognition using Artificial Neural Networks”, Proc. of IEEE International
Conference on Vehicular Electronics and Safety (ICVES), Pune, India, pp. 31 – 34, Nov.