When you obtain your frames on the telescope, the output signal is "biased" by the addition of an electronic zero-level signal of several hundreds ADU's. This level varies slightly with temperature, telescope position and who knows what else, varying mostly along the columns. Most CCD chips (but not all of them) have an overscan strip, where this additional signal is present. So, the most straight way to remove this additional signal is to fit some polynomial function to the average of several columns of overscan and to take off this fit from all the columns of the chip. Alas, not all CCD chips have overscan chips, and, moreover, there is some internal bias structure across the chip. So, usualy we take a number of zero-exposure frames (bias frames) to estimate this additional bias level.