MaskDetail.py

#!/usr/bin/python3
# -*- coding: utf-8 -*-

import vapoursynth as vs
import descale

def get_scale_offsets(scaled_w, scaled_h, origin_w, origin_h,
					  offset_l, offset_t, offset_w, offset_h):
	"""
Inverts scaling offsets (kind of like inverting the scaling itself, really).

Calculates the scale offsets necessary to scale back to its original form an
image that was scaled to the specified dimensions using the specified offsets.
	"""
	# input parameters
	scaled = [scaled_w, scaled_h]
	origin = [origin_w, origin_h]
	scaled_off = [offset_l, offset_t, offset_w, offset_h]

	# source offsets, from lengths (w / h) to crop-style (negative values)
	for i, (org_l, off_l) in enumerate(zip(origin, scaled_off[:2]), 2):
		# If the input is positive, it represents a length, rather than a crop-
		# style offset from the bottom right. Even if the output from this code
		# is positive, however, it must always be a crop-style offset, so there
		# won't be any special treatment required.
		if scaled_off[i] > 0:
			scaled_off[i] -= org_l - off_l

	# the actual scaling of the offsets
	scales = 2 * [scale / original for original, scale in zip(origin, scaled)]
	off = [-offset * scale for offset, scale in zip(scaled_off, scales * 2)]

	# target offsets, from crop-style to lengths
	for i, (scaled_l, off_l) in enumerate(zip(scaled, off[:2]), 2):
		# The input must be a crop-style offset from the bottom right, even if
		# it happens to be positive. I could process only the positive values,
		# but because I like consistency, I decided to convert all values from
		# crop-style offsets to represent a length, regardless of whether it's
		# really necessary.
		off[i] += scaled_l - off_l

	return tuple(off)

def MaskDetail(clip, final_width, final_height, RGmode=3, cutoff=None,
			   gain=0.75, expandN=2, inflateN=1, blur_more=False,
			   src_left=0, src_top=0, src_width=0, src_height=0,
			   kernel='bilinear', invkstaps=4, taps=4, mode='normal',
			   lowpasskernel='blackman', lowpassintaps=4, lowpassouttaps=3,
			   lowpassthr=None, exportlowpass=False, pclevelthr=None, b=1/3, c=1/3):
	depth = clip.format.bits_per_sample
	scale = (2 ** 16 - 1) / (2 ** depth - 1)

	if cutoff is None:
		cutoff = 17990
	else:
		cutoff *= scale

	if lowpassthr is None:
		lowpassthr = 1542
	else:
		lowpassthr *= scale

	if pclevelthr is None:
		pclevelthr = 59881
	else:
		pclevelthr *= scale

	def lowpassLut16(x):
		p = x - 0x8000
		if p > 0 and p - lowpassthr > 0:
			return x - lowpassthr
		elif p <= 0 and p + lowpassthr < 0:
			return x + lowpassthr
		else:
			return 0x8000
	def luma16(x):
		x <<= 4
		value = x & 0xFFFF
		return 0xFFFF - value if x & 0x10000 else value
	def f16(x):
		if x < cutoff:
			return 0

		result = x * gain * (0x10000 + x) / 0x10000
		return min(0xFFFF, int(result))
	def pclevelLut16(x):
		return x if x > pclevelthr else 0

	core = vs.core

	startclip = core.fmtc.bitdepth(clip, bits=16)

	original = (startclip.width, startclip.height)
	target = (final_width, final_height, src_left, src_top, src_width, src_height)

	if mode.startswith('lowpass'): # lowpass and lowpasspc
		twice = tuple(2 * o for o in original)
		lowpass = core.fmtc.resample(startclip, *twice, kernel=lowpasskernel, taps=lowpassintaps)
		lowpass = core.fmtc.resample(lowpass, *original, kernel=lowpasskernel,taps=lowpassouttaps)

		difflow = core.std.MakeDiff(startclip, lowpass, 0)
		if exportlowpass:
			return core.std.Lut(difflow, function=luma16)

		difflow = core.rgvs.RemoveGrain(difflow, mode=[1])
		difflow = core.std.Lut(difflow, function=lowpassLut16)

		startclip = core.std.MergeDiff(startclip, difflow, 0)

	if mode.startswith('pc') or mode.endswith('pc'): # pclevel and lowpasspc
		diff = core.std.Lut(startclip, function=pclevelLut16)
	else:
		temp = descale.Descale(startclip, *target[:2], kernel=kernel, taps=taps, b=b, c=c)
		
		temp = core.fmtc.resample(temp, *original, kernel=kernel, taps=taps, a1=b, a2=c)
		diff = core.std.MakeDiff(startclip, temp, 0)

	mask = core.std.Lut(diff, function=luma16).rgvs.RemoveGrain(mode=[RGmode])
	mask = core.std.Lut(mask, function=f16)

	for i in range(expandN):
		mask = core.std.Maximum(mask, planes=[0])
	for i in range(inflateN):
		mask = core.std.Inflate(mask, planes=[0])

	mask = core.fmtc.resample(mask, *target, taps=taps)
	if blur_more:
		mask = core.rgvs.RemoveGrain(mask, mode=[12,0,0])

	mask = core.std.ShufflePlanes(mask, planes=0, colorfamily=vs.GRAY)
	return core.fmtc.bitdepth(mask, bits=depth, dmode=1)