pdf_extract_kit.py 17.7 KB
Newer Older
1
from loguru import logger
myhloli's avatar
myhloli committed
2
import os
3
import time
4

5
from magic_pdf.libs.Constants import *
6
from magic_pdf.model.model_list import AtomicModel
7
8

os.environ['NO_ALBUMENTATIONS_UPDATE'] = '1'  # 禁止albumentations检查更新
myhloli's avatar
myhloli committed
9
10
11
12
13
14
try:
    import cv2
    import yaml
    import argparse
    import numpy as np
    import torch
15
    import torchtext
16

17
18
    if torchtext.__version__ >= "0.18.0":
        torchtext.disable_torchtext_deprecation_warning()
myhloli's avatar
myhloli committed
19
20
21
22
23
24
25
    from PIL import Image
    from torchvision import transforms
    from torch.utils.data import Dataset, DataLoader
    from ultralytics import YOLO
    from unimernet.common.config import Config
    import unimernet.tasks as tasks
    from unimernet.processors import load_processor
赵小蒙's avatar
update:  
赵小蒙 committed
26

27
28
except ImportError as e:
    logger.exception(e)
29
30
    logger.error(
        'Required dependency not installed, please install by \n'
31
        '"pip install magic-pdf[full] --extra-index-url https://myhloli.github.io/wheels/"')
myhloli's avatar
myhloli committed
32
    exit(1)
赵小蒙's avatar
update:  
赵小蒙 committed
33

34
from magic_pdf.model.pek_sub_modules.layoutlmv3.model_init import Layoutlmv3_Predictor
35
from magic_pdf.model.pek_sub_modules.post_process import latex_rm_whitespace
36
from magic_pdf.model.pek_sub_modules.self_modify import ModifiedPaddleOCR
37
from magic_pdf.model.pek_sub_modules.structeqtable.StructTableModel import StructTableModel
38
39
40
41
42
43
44
45
46
47
48
49
from magic_pdf.model.ppTableModel import ppTableModel


def table_model_init(table_model_type, model_path, max_time, _device_='cpu'):
    if table_model_type == STRUCT_EQTABLE:
        table_model = StructTableModel(model_path, max_time=max_time, device=_device_)
    else:
        config = {
            "model_dir": model_path,
            "device": _device_
        }
        table_model = ppTableModel(config)
50
    return table_model
51

赵小蒙's avatar
update:  
赵小蒙 committed
52

53
54
55
def mfd_model_init(weight):
    mfd_model = YOLO(weight)
    return mfd_model
赵小蒙's avatar
update:  
赵小蒙 committed
56
57


58
def mfr_model_init(weight_dir, cfg_path, _device_='cpu'):
59
60
    args = argparse.Namespace(cfg_path=cfg_path, options=None)
    cfg = Config(args)
61
    cfg.config.model.pretrained = os.path.join(weight_dir, "pytorch_model.pth")
62
63
64
65
    cfg.config.model.model_config.model_name = weight_dir
    cfg.config.model.tokenizer_config.path = weight_dir
    task = tasks.setup_task(cfg)
    model = task.build_model(cfg)
66
    model.to(_device_)
67
    model.eval()
68
    vis_processor = load_processor('formula_image_eval', cfg.config.datasets.formula_rec_eval.vis_processor.eval)
69
70
    mfr_transform = transforms.Compose([vis_processor, ])
    return [model, mfr_transform]
赵小蒙's avatar
update:  
赵小蒙 committed
71
72


73
74
75
76
77
def layout_model_init(weight, config_file, device):
    model = Layoutlmv3_Predictor(weight, config_file, device)
    return model


78
79
80
81
82
def ocr_model_init(show_log: bool = False, det_db_box_thresh=0.3, lang=None):
    if lang is not None:
        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh, lang=lang)
    else:
        model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=det_db_box_thresh)
83
84
85
    return model


86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
class MathDataset(Dataset):
    def __init__(self, image_paths, transform=None):
        self.image_paths = image_paths
        self.transform = transform

    def __len__(self):
        return len(self.image_paths)

    def __getitem__(self, idx):
        # if not pil image, then convert to pil image
        if isinstance(self.image_paths[idx], str):
            raw_image = Image.open(self.image_paths[idx])
        else:
            raw_image = self.image_paths[idx]
        if self.transform:
            image = self.transform(raw_image)
102
            return image
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
class AtomModelSingleton:
    _instance = None
    _models = {}

    def __new__(cls, *args, **kwargs):
        if cls._instance is None:
            cls._instance = super().__new__(cls)
        return cls._instance

    def get_atom_model(self, atom_model_name: str, **kwargs):
        if atom_model_name not in self._models:
            self._models[atom_model_name] = atom_model_init(model_name=atom_model_name, **kwargs)
        return self._models[atom_model_name]


def atom_model_init(model_name: str, **kwargs):

    if model_name == AtomicModel.Layout:
        atom_model = layout_model_init(
            kwargs.get("layout_weights"),
            kwargs.get("layout_config_file"),
            kwargs.get("device")
        )
    elif model_name == AtomicModel.MFD:
        atom_model = mfd_model_init(
            kwargs.get("mfd_weights")
        )
    elif model_name == AtomicModel.MFR:
        atom_model = mfr_model_init(
            kwargs.get("mfr_weight_dir"),
            kwargs.get("mfr_cfg_path"),
            kwargs.get("device")
        )
    elif model_name == AtomicModel.OCR:
        atom_model = ocr_model_init(
            kwargs.get("ocr_show_log"),
141
142
            kwargs.get("det_db_box_thresh"),
            kwargs.get("lang")
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
        )
    elif model_name == AtomicModel.Table:
        atom_model = table_model_init(
            kwargs.get("table_model_type"),
            kwargs.get("table_model_path"),
            kwargs.get("table_max_time"),
            kwargs.get("device")
        )
    else:
        logger.error("model name not allow")
        exit(1)

    return atom_model


158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
#  Unified crop img logic
def crop_img(input_res, input_pil_img, crop_paste_x=0, crop_paste_y=0):
    crop_xmin, crop_ymin = int(input_res['poly'][0]), int(input_res['poly'][1])
    crop_xmax, crop_ymax = int(input_res['poly'][4]), int(input_res['poly'][5])
    # Create a white background with an additional width and height of 50
    crop_new_width = crop_xmax - crop_xmin + crop_paste_x * 2
    crop_new_height = crop_ymax - crop_ymin + crop_paste_y * 2
    return_image = Image.new('RGB', (crop_new_width, crop_new_height), 'white')

    # Crop image
    crop_box = (crop_xmin, crop_ymin, crop_xmax, crop_ymax)
    cropped_img = input_pil_img.crop(crop_box)
    return_image.paste(cropped_img, (crop_paste_x, crop_paste_y))
    return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width, crop_new_height]
    return return_image, return_list


175
class CustomPEKModel:
176

177
178
179
180
181
182
183
184
185
186
187
188
189
190
    def __init__(self, ocr: bool = False, show_log: bool = False, **kwargs):
        """
        ======== model init ========
        """
        # 获取当前文件(即 pdf_extract_kit.py)的绝对路径
        current_file_path = os.path.abspath(__file__)
        # 获取当前文件所在的目录(model)
        current_dir = os.path.dirname(current_file_path)
        # 上一级目录(magic_pdf)
        root_dir = os.path.dirname(current_dir)
        # model_config目录
        model_config_dir = os.path.join(root_dir, 'resources', 'model_config')
        # 构建 model_configs.yaml 文件的完整路径
        config_path = os.path.join(model_config_dir, 'model_configs.yaml')
191
        with open(config_path, "r", encoding='utf-8') as f:
192
193
194
195
            self.configs = yaml.load(f, Loader=yaml.FullLoader)
        # 初始化解析配置
        self.apply_layout = kwargs.get("apply_layout", self.configs["config"]["layout"])
        self.apply_formula = kwargs.get("apply_formula", self.configs["config"]["formula"])
196
        # table config
197
        self.table_config = kwargs.get("table_config", self.configs["config"]["table_config"])
198
        self.apply_table = self.table_config.get("is_table_recog_enable", False)
199
        self.table_max_time = self.table_config.get("max_time", TABLE_MAX_TIME_VALUE)
200
        self.table_model_type = self.table_config.get("model", TABLE_MASTER)
201
        self.apply_ocr = ocr
202
        self.lang = kwargs.get("lang", None)
203
        logger.info(
204
205
            "DocAnalysis init, this may take some times. apply_layout: {}, apply_formula: {}, apply_ocr: {}, apply_table: {}, lang: {}".format(
                self.apply_layout, self.apply_formula, self.apply_ocr, self.apply_table, self.lang
赵小蒙's avatar
update:  
赵小蒙 committed
206
            )
207
208
209
        )
        assert self.apply_layout, "DocAnalysis must contain layout model."
        # 初始化解析方案
210
        self.device = kwargs.get("device", self.configs["config"]["device"])
211
        logger.info("using device: {}".format(self.device))
212
        models_dir = kwargs.get("models_dir", os.path.join(root_dir, "resources", "models"))
213
        logger.info("using models_dir: {}".format(models_dir))
214

215
216
        atom_model_manager = AtomModelSingleton()

217
218
219
        # 初始化公式识别
        if self.apply_formula:
            # 初始化公式检测模型
220
221
222
223
224
            # self.mfd_model = mfd_model_init(str(os.path.join(models_dir, self.configs["weights"]["mfd"])))
            self.mfd_model = atom_model_manager.get_atom_model(
                atom_model_name=AtomicModel.MFD,
                mfd_weights=str(os.path.join(models_dir, self.configs["weights"]["mfd"]))
            )
225
            # 初始化公式解析模型
226
227
            mfr_weight_dir = str(os.path.join(models_dir, self.configs["weights"]["mfr"]))
            mfr_cfg_path = str(os.path.join(model_config_dir, "UniMERNet", "demo.yaml"))
228
229
230
231
232
233
234
235
            # self.mfr_model, mfr_vis_processors = mfr_model_init(mfr_weight_dir, mfr_cfg_path, _device_=self.device)
            # self.mfr_transform = transforms.Compose([mfr_vis_processors, ])
            self.mfr_model, self.mfr_transform = atom_model_manager.get_atom_model(
                atom_model_name=AtomicModel.MFR,
                mfr_weight_dir=mfr_weight_dir,
                mfr_cfg_path=mfr_cfg_path,
                device=self.device
            )
236
237

        # 初始化layout模型
238
239
240
241
242
243
244
245
246
        # self.layout_model = Layoutlmv3_Predictor(
        #     str(os.path.join(models_dir, self.configs['weights']['layout'])),
        #     str(os.path.join(model_config_dir, "layoutlmv3", "layoutlmv3_base_inference.yaml")),
        #     device=self.device
        # )
        self.layout_model = atom_model_manager.get_atom_model(
            atom_model_name=AtomicModel.Layout,
            layout_weights=str(os.path.join(models_dir, self.configs['weights']['layout'])),
            layout_config_file=str(os.path.join(model_config_dir, "layoutlmv3", "layoutlmv3_base_inference.yaml")),
247
248
            device=self.device
        )
249
250
        # 初始化ocr
        if self.apply_ocr:
drunkpig's avatar
drunkpig committed
251

252
253
254
255
            # self.ocr_model = ModifiedPaddleOCR(show_log=show_log, det_db_box_thresh=0.3)
            self.ocr_model = atom_model_manager.get_atom_model(
                atom_model_name=AtomicModel.OCR,
                ocr_show_log=show_log,
256
257
                det_db_box_thresh=0.3,
                lang=self.lang
258
            )
259
        # init table model
260
        if self.apply_table:
261
            table_model_dir = self.configs["weights"][self.table_model_type]
262
263
264
265
266
267
268
269
270
            # self.table_model = table_model_init(self.table_model_type, str(os.path.join(models_dir, table_model_dir)),
            #                                     max_time=self.table_max_time, _device_=self.device)
            self.table_model = atom_model_manager.get_atom_model(
                atom_model_name=AtomicModel.Table,
                table_model_type=self.table_model_type,
                table_model_path=str(os.path.join(models_dir, table_model_dir)),
                table_max_time=self.table_max_time,
                device=self.device
            )
drunkpig's avatar
drunkpig committed
271

272
        logger.info('DocAnalysis init done!')
赵小蒙's avatar
update:  
赵小蒙 committed
273

274
275
    def __call__(self, image):

276
277
278
        latex_filling_list = []
        mf_image_list = []

279
280
281
282
283
284
        # layout检测
        layout_start = time.time()
        layout_res = self.layout_model(image, ignore_catids=[])
        layout_cost = round(time.time() - layout_start, 2)
        logger.info(f"layout detection cost: {layout_cost}")

285
286
        pil_img = Image.fromarray(image)

287
288
289
290
291
292
293
294
295
296
297
298
299
        if self.apply_formula:
            # 公式检测
            mfd_res = self.mfd_model.predict(image, imgsz=1888, conf=0.25, iou=0.45, verbose=True)[0]
            for xyxy, conf, cla in zip(mfd_res.boxes.xyxy.cpu(), mfd_res.boxes.conf.cpu(), mfd_res.boxes.cls.cpu()):
                xmin, ymin, xmax, ymax = [int(p.item()) for p in xyxy]
                new_item = {
                    'category_id': 13 + int(cla.item()),
                    'poly': [xmin, ymin, xmax, ymin, xmax, ymax, xmin, ymax],
                    'score': round(float(conf.item()), 2),
                    'latex': '',
                }
                layout_res.append(new_item)
                latex_filling_list.append(new_item)
300
301
                # bbox_img = get_croped_image(pil_img, [xmin, ymin, xmax, ymax])
                bbox_img = pil_img.crop((xmin, ymin, xmax, ymax))
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
                mf_image_list.append(bbox_img)

            # 公式识别
            mfr_start = time.time()
            dataset = MathDataset(mf_image_list, transform=self.mfr_transform)
            dataloader = DataLoader(dataset, batch_size=64, num_workers=0)
            mfr_res = []
            for mf_img in dataloader:
                mf_img = mf_img.to(self.device)
                output = self.mfr_model.generate({'image': mf_img})
                mfr_res.extend(output['pred_str'])
            for res, latex in zip(latex_filling_list, mfr_res):
                res['latex'] = latex_rm_whitespace(latex)
            mfr_cost = round(time.time() - mfr_start, 2)
            logger.info(f"formula nums: {len(mf_image_list)}, mfr time: {mfr_cost}")
317

318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
        # Select regions for OCR / formula regions / table regions
        ocr_res_list = []
        table_res_list = []
        single_page_mfdetrec_res = []
        for res in layout_res:
            if int(res['category_id']) in [13, 14]:
                single_page_mfdetrec_res.append({
                    "bbox": [int(res['poly'][0]), int(res['poly'][1]),
                             int(res['poly'][4]), int(res['poly'][5])],
                })
            elif int(res['category_id']) in [0, 1, 2, 4, 6, 7]:
                ocr_res_list.append(res)
            elif int(res['category_id']) in [5]:
                table_res_list.append(res)

myhloli's avatar
myhloli committed
333
        # ocr识别
334
        if self.apply_ocr:
335
            ocr_start = time.time()
336
            # Process each area that requires OCR processing
337
            for res in ocr_res_list:
338
339
340
                new_image, useful_list = crop_img(res, pil_img, crop_paste_x=50, crop_paste_y=50)
                paste_x, paste_y, xmin, ymin, xmax, ymax, new_width, new_height = useful_list
                # Adjust the coordinates of the formula area
341
342
343
                adjusted_mfdetrec_res = []
                for mf_res in single_page_mfdetrec_res:
                    mf_xmin, mf_ymin, mf_xmax, mf_ymax = mf_res["bbox"]
344
                    # Adjust the coordinates of the formula area to the coordinates relative to the cropping area
345
346
347
348
                    x0 = mf_xmin - xmin + paste_x
                    y0 = mf_ymin - ymin + paste_y
                    x1 = mf_xmax - xmin + paste_x
                    y1 = mf_ymax - ymin + paste_y
349
                    # Filter formula blocks outside the graph
350
                    if any([x1 < 0, y1 < 0]) or any([x0 > new_width, y0 > new_height]):
351
352
353
354
355
356
                        continue
                    else:
                        adjusted_mfdetrec_res.append({
                            "bbox": [x0, y0, x1, y1],
                        })

357
                # OCR recognition
358
359
                new_image = cv2.cvtColor(np.asarray(new_image), cv2.COLOR_RGB2BGR)
                ocr_res = self.ocr_model.ocr(new_image, mfd_res=adjusted_mfdetrec_res)[0]
360

361
                # Integration results
362
363
364
365
366
                if ocr_res:
                    for box_ocr_res in ocr_res:
                        p1, p2, p3, p4 = box_ocr_res[0]
                        text, score = box_ocr_res[1]

367
                        # Convert the coordinates back to the original coordinate system
368
369
370
371
372
373
374
375
376
377
378
379
                        p1 = [p1[0] - paste_x + xmin, p1[1] - paste_y + ymin]
                        p2 = [p2[0] - paste_x + xmin, p2[1] - paste_y + ymin]
                        p3 = [p3[0] - paste_x + xmin, p3[1] - paste_y + ymin]
                        p4 = [p4[0] - paste_x + xmin, p4[1] - paste_y + ymin]

                        layout_res.append({
                            'category_id': 15,
                            'poly': p1 + p2 + p3 + p4,
                            'score': round(score, 2),
                            'text': text,
                        })

380
381
382
            ocr_cost = round(time.time() - ocr_start, 2)
            logger.info(f"ocr cost: {ocr_cost}")

383
384
        # 表格识别 table recognition
        if self.apply_table:
385
386
387
388
389
            table_start = time.time()
            for res in table_res_list:
                new_image, _ = crop_img(res, pil_img)
                single_table_start_time = time.time()
                logger.info("------------------table recognition processing begins-----------------")
390
391
                latex_code = None
                html_code = None
392
393
                if self.table_model_type == STRUCT_EQTABLE:
                    with torch.no_grad():
394
                        latex_code = self.table_model.image2latex(new_image)[0]
395
396
                else:
                    html_code = self.table_model.img2html(new_image)
drunkpig's avatar
drunkpig committed
397

398
399
400
401
402
                run_time = time.time() - single_table_start_time
                logger.info(f"------------table recognition processing ends within {run_time}s-----")
                if run_time > self.table_max_time:
                    logger.warning(f"------------table recognition processing exceeds max time {self.table_max_time}s----------")
                # 判断是否返回正常
403
404
405
406
407
408
409
410
411
412

                if latex_code:
                    expected_ending = latex_code.strip().endswith('end{tabular}') or latex_code.strip().endswith(
                        'end{table}')
                    if expected_ending:
                        res["latex"] = latex_code
                    else:
                        logger.warning(f"------------table recognition processing fails----------")
                elif html_code:
                    res["html"] = html_code
413
414
415
416
417
                else:
                    logger.warning(f"------------table recognition processing fails----------")
            table_cost = round(time.time() - table_start, 2)
            logger.info(f"table cost: {table_cost}")

418
        return layout_res