File size: 5,116 Bytes
2b7aae2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
"""
Image processing utilities for ML services.
"""

import io
import math
import base64
import numpy as np
import cv2
import PIL.Image


MARGIN_DIVIDER = 8


def array_from_image_stream(buffer):
	"""Convert image bytes to numpy array (H, W, C)."""
	if not isinstance(buffer, bytes):
		return None

	stream = io.BytesIO(buffer)
	image = PIL.Image.open(stream)

	arr = np.array(image)
	if len(arr.shape) >= 3:
		return arr[:, :, :3]  # RGB only

	arr = np.expand_dims(arr, -1)
	return arr


def slice_feature(source, width, overlapping=0.25, padding=False):
	"""
	Slice image horizontally into overlapping pieces.
	source: (height, width, channel)
	yields: (height, width, channel) pieces
	"""
	step = math.floor(width - source.shape[0] * overlapping)

	if padding:
		margin = math.floor(source.shape[0] * overlapping) // 2
		center = source
		source = np.zeros((source.shape[0], source.shape[1] + margin * 2, source.shape[2]))
		source[:, margin:-margin, :] = center
		source[:, :margin, :] = center[:, :1, :]
		source[:, -margin:, :] = center[:, -1:, :]

	for x in range(0, source.shape[1], step):
		if x + width <= source.shape[1]:
			yield source[:, x:x + width, :]
		else:
			sliced = np.ones((source.shape[0], width, source.shape[2]), dtype=np.float32) * 255
			sliced[:, :source.shape[1] - x, :] = source[:, x:, :]
			yield sliced


def splice_pieces(pieces, margin_divider, keep_margin=False):
	"""
	Splice output tensor pieces back together.
	pieces: (batch, channel, height, width)
	returns: (channel, height, width)
	"""
	piece_height, piece_width = pieces.shape[2:]
	margin_width = piece_height // margin_divider
	patch_width = piece_width - margin_width * 2
	result = np.zeros((pieces.shape[1], pieces.shape[2], patch_width * pieces.shape[0]), dtype=np.float32)

	for i, piece in enumerate(pieces):
		result[:, :, i * patch_width : (i + 1) * patch_width] = piece[:, :, margin_width:-margin_width]

	if keep_margin:
		return np.concatenate((pieces[0, :, :, :margin_width], result, pieces[-1, :, :, -margin_width:]), axis=2)

	return result


def soft_splice_pieces(pieces, margin_divider):
	"""
	Splice pieces with soft blending at overlaps.
	pieces: (batch, channel, height, width)
	returns: (channel, height, width)
	"""
	batches, channels, piece_height, piece_width = pieces.shape
	overlap_width = piece_height * 2 // margin_divider
	segment_width = piece_width - overlap_width

	slope = np.arange(overlap_width, dtype=np.float32) / overlap_width
	slope = slope.reshape((1, 1, overlap_width))
	inv_slope = 1 - slope

	result = np.zeros((channels, piece_height, segment_width * batches + overlap_width), dtype=np.float32)

	for i, piece in enumerate(pieces):
		if i > 0:
			piece[:, :, :overlap_width] *= slope

		if i < batches - 1:
			piece[:, :, -overlap_width:] *= inv_slope

		result[:, :, segment_width * i:segment_width * i + piece_width] += piece

	return result


def splice_output_tensor(tensor, keep_margin=False, soft=False, margin_divider=MARGIN_DIVIDER):
	"""
	Splice PyTorch tensor output back together.
	tensor: PyTorch tensor or numpy array
	"""
	if tensor is None:
		return None

	arr = tensor.cpu().numpy() if hasattr(tensor, 'cpu') else tensor

	if soft:
		return soft_splice_pieces(arr, margin_divider)

	return splice_pieces(arr, margin_divider, keep_margin=keep_margin)


def mask_to_alpha(mask, frac_y=False):
	"""
	Convert mask to LA (grayscale + alpha) format.
	mask: [fore(h, w), back(h, w)]
	"""
	fore = mask[0] * 255
	fore = np.stack([np.zeros(mask[0].shape, np.float32), fore], axis=2)
	fore = np.uint8(np.clip(fore, 0, 255))
	return fore


def gauge_to_rgb(gauge, frac_y=False):
	"""
	Convert gauge map to RGB.
	gauge: [Y(h, w), K(h, w)]
	"""
	mapy = gauge[0] * 8 + 128
	mapk = gauge[1] * 127 + 128

	if frac_y:
		B, R = np.modf(mapy)
		result = np.stack([B * 256, mapk, R], axis=2)
	else:
		result = np.stack([np.zeros(mapy.shape, np.float32), mapk, mapy], axis=2)

	return np.uint8(np.clip(result, 0, 255))


def encode_image_bytes(image, ext='.png', quality=80):
	"""Encode PIL Image to bytes."""
	fp = io.BytesIO()
	image.save(fp, PIL.Image.registered_extensions()[ext], quality=quality)
	return fp.getvalue()


def encode_image_base64(image, ext='.png', quality=80):
	"""Encode PIL Image to base64 data URI."""
	image_bytes = encode_image_bytes(image, ext, quality=quality)
	b64 = base64.b64encode(image_bytes)
	fmt = ext.replace('.', '')
	return f'data:image/{fmt};base64,' + b64.decode('ascii')


def resize_page_image(img, size):
	"""Resize page image maintaining aspect ratio, padding with white."""
	w, h = size
	filled_height = img.shape[0] * w // img.shape[1]
	img = cv2.resize(img, (w, filled_height), interpolation=cv2.INTER_AREA)
	if len(img.shape) < 3:
		img = np.expand_dims(img, -1)

	if filled_height < h:
		result = np.ones((h, w, img.shape[2]), dtype=np.uint8) * 255
		result[:filled_height] = img
		return result

	return img[:h]


def normalize_image_dimension(image):
	"""Ensure image dimensions are divisible by 4 (for UNet)."""
	n, h, w, c = image.shape
	if (h % 4 != 0) | (w % 4 != 0):
		return image[:, :h - h % 4, :w - w % 4, :]
	return image