File size: 12,598 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
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
"""
Layout prediction service.
Detects page layout: staff boxes, lines, intervals, rotation angle.
Includes system/staff detection for the full prediction pipeline.
"""

import numpy as np
import torch
import cv2
import PIL.Image
import hashlib
import io
import logging

from predictors.torchscript_predictor import TorchScriptPredictor
from common.image_utils import (
	array_from_image_stream, resize_page_image, normalize_image_dimension,
	encode_image_base64
)
from common.transform import Composer


RESIZE_WIDTH = 600
CANVAS_WIDTH_MIN = 1024

SYSTEM_HEIGHT_ENLARGE = 0.02
SYSTEM_LEFT_ENLARGE = 0.03
SYSTEM_RIGHT_ENLARGE = 0.01

STAFF_PADDING_LEFT = 32
STAFF_HEIGHT_UNITS = 24
UNIT_SIZE = 8


def _detect_systems(image):
	"""Detect musical systems (staff groups) from max-channel heatmap."""
	height, width = image.shape

	blur = cv2.GaussianBlur(image, (5, 5), 0)
	thresh = cv2.adaptiveThreshold(blur, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 201, -40)
	contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

	marginLeft = SYSTEM_LEFT_ENLARGE * width
	marginRight = SYSTEM_RIGHT_ENLARGE * width

	areas = []
	for contour in contours:
		x, y, w, h = cv2.boundingRect(contour)
		rw = w / width
		rh = h / width
		if (rw > 0.6 and rh > 0.02) or (rw > 0.12 and rh > 0.2):
			left = max(x - marginLeft, 0)
			right = min(x + w + marginRight, width)
			areas.append({'x': left, 'y': y, 'width': right - left, 'height': h})
	areas.sort(key=lambda a: a['y'])

	# Enlarge heights to include surrounding space
	marginY = SYSTEM_HEIGHT_ENLARGE * width
	marginYMax = marginY * 8
	ctx = {'lastMargin': 0}

	def enlarge(i, area, ctx):
		top = area['y']
		bottom = top + area['height']
		if i > 0:
			lastArea = areas[i - 1]
			ctx['lastMargin'] = max(ctx['lastMargin'], lastArea['y'] + lastArea['height'], top - marginYMax)
			top = max(0, min(top - marginY, ctx['lastMargin']))
		else:
			top = min(top, max(marginY, top - marginYMax))
		if i < len(areas) - 1:
			nextArea = areas[i + 1]
			bottom = min(height, max(bottom + marginY, nextArea['y']), bottom + marginYMax)
		else:
			bottom = min(height, bottom + marginYMax)
		return {'top': top, 'bottom': bottom}

	enlarges = [enlarge(i, area, ctx) for i, area in enumerate(areas)]
	for i, area in enumerate(areas):
		area['y'] = enlarges[i]['top']
		area['height'] = enlarges[i]['bottom'] - enlarges[i]['top']

	return {'areas': areas}


def _detect_staves_from_hbl(HB, HL, interval):
	"""Detect individual staves within a system using staff-box and horizontal-line heatmaps."""
	_, HB = cv2.threshold(HB, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
	contours, _ = cv2.findContours(HB, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

	height, width = HB.shape
	STAFF_HEIGHT_MIN = interval * 3
	STAFF_WIDTH_MIN = max(width * 0.6, width - interval * 12)
	UPSCALE = 4
	upInterval = interval * UPSCALE

	rects = map(cv2.boundingRect, contours)
	rects = filter(lambda rect: rect[2] > STAFF_WIDTH_MIN and rect[3] > STAFF_HEIGHT_MIN, rects)
	rects = sorted(rects, key=lambda rect: rect[1])

	# Merge overlapping rectangles
	preRects = []
	for rect in rects:
		x, y, w, h = rect
		ri = next((i for i, rc in enumerate(preRects)
			if (y + h / 2) - (rc[1] + rc[3] / 2) < (h + rc[3]) / 2), -1)
		if ri < 0:
			preRects.append(rect)
		else:
			rc = list(preRects[ri])
			if w > rc[2]:
				preRects[ri] = (min(x, rc[0]), rect[1], max(x + w, rc[0] + rc[2]) - x, rect[3])
			else:
				rc[0] = min(x, rc[0])
				rc[2] = max(x + w, rc[0] + rc[2]) - rc[0]

	if len(preRects) == 0:
		return {'reason': 'no block rectangles detected'}

	phi1 = min(rc[0] for rc in preRects)
	phi2 = min(rc[0] + rc[2] for rc in preRects)

	middleRhos = []
	for rect in preRects:
		rx, ry, rw, rh = rect
		x = max(rx, 0)
		y = max(round(ry - interval), 0)
		roi = (x, y, min(rw, width - x), min(round(rh + interval + ry - y), height - y))
		staffLines = HL[roi[1]:roi[1] + roi[3], roi[0]:roi[0] + roi[2]]

		hlineColumn = cv2.resize(staffLines, (1, staffLines.shape[0] * UPSCALE), 0, 0, cv2.INTER_LINEAR).flatten()
		i1 = round(upInterval)
		kernel = np.zeros(i1 * 4 + 1)
		kernel[::i1] = 1
		convolutionLine = np.convolve(hlineColumn, kernel)
		convolutionLineMax = np.max(convolutionLine)
		middleY = np.where(convolutionLine == convolutionLineMax)[0][0] - round(upInterval * 2)
		middleRhos.append(y + middleY / UPSCALE)

	return {
		'interval': interval,
		'phi1': phi1,
		'phi2': phi2,
		'middleRhos': middleRhos,
	}


def _scale_detection(detection, scale):
	"""Scale detection coordinates by given factor."""
	areas = [{
		'x': a['x'] * scale,
		'y': a['y'] * scale,
		'width': a['width'] * scale,
		'height': a['height'] * scale,
		'staff_images': a.get('staff_images'),
		'staves': {
			'interval': a['staves']['interval'] * scale,
			'phi1': a['staves']['phi1'] * scale,
			'phi2': a['staves']['phi2'] * scale,
			'middleRhos': [rho * scale for rho in a['staves']['middleRhos']],
		} if (a.get('staves') and a['staves'].get('middleRhos') is not None) else None,
	} for a in detection['areas']]
	return {'areas': areas}


class PageLayout:
	"""Process layout heatmap to extract page parameters."""

	def __init__(self, heatmap):
		"""
		heatmap: (3, H, W) - channels: [VL, StaffBox, HL]
		"""
		lines_map = heatmap[2]
		self.interval = self.measure_interval(lines_map)

		heatmap_uint8 = np.uint8(heatmap * 255)
		heatmap_uint8 = np.moveaxis(heatmap_uint8, 0, -1)
		self.image = PIL.Image.fromarray(heatmap_uint8, 'RGB')
		self.heatmap = heatmap_uint8

		staves_map = heatmap_uint8[:, :, 1]
		self.theta = self.measure_theta(staves_map)

	def json(self):
		return {
			'image': encode_image_base64(self.image),
			'theta': self.theta,
			'interval': self.interval,
		}

	def detect(self, image, ratio):
		"""Full detection: systems, staves, staff images."""
		if self.theta is None:
			return {
				'theta': self.theta,
				'interval': self.interval * image.shape[1] / RESIZE_WIDTH,
				'detection': None,
			}

		original_size = (image.shape[1], image.shape[0])
		aligned_height = int(image.shape[1] * ratio)
		if image.shape[0] < aligned_height:
			image = np.pad(image, ((0, aligned_height - image.shape[0]), (0, 0), (0, 0)), mode='constant')
		elif image.shape[0] > aligned_height:
			image = image[:aligned_height]

		canvas_size = (original_size[0], aligned_height)
		while canvas_size[0] < CANVAS_WIDTH_MIN:
			canvas_size = (canvas_size[0] * 2, canvas_size[1] * 2)
		if canvas_size[0] > original_size[0]:
			image = cv2.resize(image, canvas_size)

		rot_mat = cv2.getRotationMatrix2D((canvas_size[0] / 2, canvas_size[1] / 2), self.theta * 180 / np.pi, 1)
		image = cv2.warpAffine(image, rot_mat, canvas_size, flags=cv2.INTER_CUBIC)
		if len(image.shape) < 3:
			image = np.expand_dims(image, -1)

		heatmap = cv2.resize(self.heatmap, (canvas_size[0], round(canvas_size[0] * self.heatmap.shape[0] / self.heatmap.shape[1])), interpolation=cv2.INTER_CUBIC)
		if heatmap.shape[0] > canvas_size[1]:
			heatmap = heatmap[:canvas_size[1]]
		elif heatmap.shape[0] < canvas_size[1]:
			heatmap = np.pad(heatmap, ((0, canvas_size[1] - heatmap.shape[0]), (0, 0), (0, 0)), mode='constant')
		heatmap = cv2.warpAffine(heatmap, rot_mat, canvas_size, flags=cv2.INTER_LINEAR)

		HB = heatmap[:, :, 1]
		HL = heatmap[:, :, 2]
		block = heatmap.max(axis=2)
		detection = _detect_systems(block)

		canvas_interval = self.interval * canvas_size[0] / RESIZE_WIDTH

		for si, area in enumerate(detection['areas']):
			l, r, t, b = map(round, (area['x'], area['x'] + area['width'], area['y'], area['y'] + area['height']))
			system_image = image[t:b, l:r, :]
			hb = HB[t:b, l:r]
			hl = HL[t:b, l:r]

			area['staves'] = _detect_staves_from_hbl(hb, hl, canvas_interval)

			if not area.get('staves') or area['staves'].get('middleRhos') is None:
				continue

			area['staff_images'] = []
			interval = area['staves']['interval']
			unit_scaling = UNIT_SIZE / interval
			padding_left = round(STAFF_PADDING_LEFT * UNIT_SIZE / interval / unit_scaling)
			staff_width = round(system_image.shape[1] * unit_scaling) + STAFF_PADDING_LEFT
			staff_size = (staff_width, STAFF_HEIGHT_UNITS * UNIT_SIZE)

			for ssi, rho in enumerate(area['staves']['middleRhos']):
				map_x = (np.tile(np.arange(staff_size[0], dtype=np.float32), (staff_size[1], 1)) - STAFF_PADDING_LEFT) / unit_scaling
				map_y = (np.tile(np.arange(staff_size[1], dtype=np.float32), (staff_size[0], 1)).T - staff_size[1] / 2) / unit_scaling + rho
				map_x, map_y = map_x.astype(np.float32), map_y.astype(np.float32)

				staff_image = cv2.remap(system_image, map_x, map_y, cv2.INTER_CUBIC, borderMode=cv2.BORDER_CONSTANT, borderValue=(255, 255, 255))

				# Encode staff image as PNG bytes for downstream predictors
				_, png_data = cv2.imencode('.png', staff_image)
				staff_bytes = png_data.tobytes()

				area['staff_images'].append({
					'hash': None,
					'image': staff_bytes,
					'position': {
						'x': -STAFF_PADDING_LEFT / UNIT_SIZE,
						'y': -STAFF_HEIGHT_UNITS / 2,
						'width': staff_size[0] / UNIT_SIZE,
						'height': staff_size[1] / UNIT_SIZE,
					},
				})

		page_interval = self.interval * original_size[0] / RESIZE_WIDTH
		return {
			'sourceSize': {
				'width': original_size[0],
				'height': original_size[1],
			},
			'theta': self.theta,
			'interval': page_interval,
			'detection': _scale_detection(detection, original_size[0] / canvas_size[0]),
		}

	@staticmethod
	def measure_theta(heatmap):
		"""Measure page rotation angle using Hough lines."""
		edges = cv2.Canny(heatmap, 50, 150, apertureSize=3)
		lines = cv2.HoughLines(
			edges, 1, np.pi / 18000,
			round(heatmap.shape[1] * 0.4),
			min_theta=np.pi * 0.48,
			max_theta=np.pi * 0.52
		)
		if lines is None:
			return None

		avg_theta = sum(line[0][1] for line in lines) / len(lines)
		return float(avg_theta - np.pi / 2)

	@staticmethod
	def measure_interval(heatmap):
		"""Measure staff line interval using autocorrelation."""
		UPSCALE = 4
		width = heatmap.shape[1]

		heatmap = cv2.resize(
			heatmap,
			(heatmap.shape[1] // UPSCALE, heatmap.shape[0] * UPSCALE),
			interpolation=cv2.INTER_LINEAR
		)

		interval_min = round(width * 0.002 * UPSCALE)
		interval_max = round(width * 0.025 * UPSCALE)

		brights = []
		for y in range(interval_min, interval_max):
			m1, m2 = heatmap[y:], heatmap[:-y]
			p = np.multiply(m1, m2)
			brights.append(np.mean(p))

		# Subtract 2x interval to weaken harmonics
		brights = np.array([brights])
		brights2 = cv2.resize(brights, (brights.shape[1] * 2, 1))

		brights = brights.flatten()[interval_min:]
		brights2 = brights2.flatten()[:len(brights)]

		brights -= brights2 * 0.5

		return (interval_min * 2 + int(np.argmax(brights))) / UPSCALE


class LayoutService(TorchScriptPredictor):
	"""Layout prediction service using TorchScript model."""

	# Default transform pipeline
	DEFAULT_TRANS = ['Mono', 'HWC2CHW']

	def __init__(self, model_path, device='cuda', trans=None, **kwargs):
		super().__init__(model_path, device)
		self.composer = Composer(trans or self.DEFAULT_TRANS)

	def predict(self, streams, **kwargs):
		"""
		Predict page layout from image streams (basic mode).
		streams: list of image byte buffers
		yields: layout JSON results
		"""
		for stream in streams:
			image = array_from_image_stream(stream)
			if image is None:
				yield {'error': 'Invalid image'}
				continue

			image = np.expand_dims(image, 0)  # (1, H, W, C)
			image = normalize_image_dimension(image)

			batch, _ = self.composer(image, np.ones((1, 4, 4, 2)))
			batch = torch.from_numpy(batch).to(self.device)

			output = self.run_inference(batch)
			output = output.cpu().numpy()
			hotmap = output[0]  # (C, H, W)

			yield PageLayout(hotmap).json()

	def predictDetection(self, streams):
		"""
		Predict layout with full system/staff detection.
		streams: list of image byte buffers
		yields: detection results with sourceSize, theta, interval, detection.areas
		"""
		images = [array_from_image_stream(stream) for stream in streams]

		ratio = max(img.shape[0] / img.shape[1] for img in images)
		height = int(RESIZE_WIDTH * ratio)
		height += -height % 4
		unified_images = [resize_page_image(img, (RESIZE_WIDTH, height)) for img in images]
		image_array = np.stack(unified_images, axis=0)

		batch, _ = self.composer(image_array, np.ones((1, 4, 4, 2)))
		batch = torch.from_numpy(batch).to(self.device)

		with torch.no_grad():
			output = self.run_inference(batch)
			output = output.cpu().numpy()

			for i, heatmap in enumerate(output):
				image = images[i]
				layout = PageLayout(heatmap)
				result = layout.detect(image, ratio)
				yield result