Input values are non-linearly digitally filtered to produce as an output value the R.sup.th-largest value among the K input values, wherein the R.sup.th-largest value may be the median value with R=(K+1)/2. The input values are provided in any binary fixed point number representation. A respective selected bit of all the input values is evaluated to determine the R.sup.th-largest bit value of this bit among the K input values, and this gives the value of the corresponding bit of the output value. In the input values of which the selected bit does not correspond to the determined R.sup.th-largest bit value, all of the subsequent less-significant bits are set as dummy bits equal to the selected bit, which excludes these input values from being the R.sup.th-largest value in the subsequent evaluation of the successive less-significant bits. Then these steps are repeated for the next selected less-significant bit, and so forth, thereby respectively determining the successive bits of the output value. Minor modifications are necessary depending on the particular numerical representation. The method can be used for implementing digital filters, in particular for signal processing in systems for object recognition.