Detect words using maxima instead of mean

[kolhizin]

Made major change in algorithm before implementing MVP 06: mean -> max. It should increase quality of detection.

Test:

• rolling mean
• filter segment whether it is too small
• filter segment whether inside mean value is too small

[kolhizin]

Works far better, but still remains principal issue:

• joins words, when should not (e.g. "creeping" from characters from previous or next line)
• separates words, when should not (e.g. capital P in russian, i.e. part of the symbol is going by the edge of the line)