extract_textpoint_fast.py 16.9 KB
Newer Older
1
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
J
Jethong 已提交
2 3 4 5 6 7 8 9 10 11 12 13
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
J
Jethong 已提交
14 15 16 17 18 19
"""Contains various CTC decoders."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import cv2
J
Jethong 已提交
20 21
import math

J
Jethong 已提交
22 23
import numpy as np
from itertools import groupby
J
Jethong 已提交
24
from cv2.ximgproc import thinning as thin
J
Jethong 已提交
25 26


J
Jethong 已提交
27 28 29 30 31 32 33 34 35 36 37
def get_dict(character_dict_path):
    character_str = ""
    with open(character_dict_path, "rb") as fin:
        lines = fin.readlines()
        for line in lines:
            line = line.decode('utf-8').strip("\n").strip("\r\n")
            character_str += line
        dict_character = list(character_str)
    return dict_character


J
Jethong 已提交
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
def softmax(logits):
    """
    logits: N x d
    """
    max_value = np.max(logits, axis=1, keepdims=True)
    exp = np.exp(logits - max_value)
    exp_sum = np.sum(exp, axis=1, keepdims=True)
    dist = exp / exp_sum
    return dist


def get_keep_pos_idxs(labels, remove_blank=None):
    """
    Remove duplicate and get pos idxs of keep items.
    The value of keep_blank should be [None, 95].
    """
    duplicate_len_list = []
    keep_pos_idx_list = []
    keep_char_idx_list = []
    for k, v_ in groupby(labels):
        current_len = len(list(v_))
        if k != remove_blank:
            current_idx = int(sum(duplicate_len_list) + current_len // 2)
            keep_pos_idx_list.append(current_idx)
            keep_char_idx_list.append(k)
        duplicate_len_list.append(current_len)
    return keep_char_idx_list, keep_pos_idx_list


def remove_blank(labels, blank=0):
    new_labels = [x for x in labels if x != blank]
    return new_labels


def insert_blank(labels, blank=0):
    new_labels = [blank]
    for l in labels:
        new_labels += [l, blank]
    return new_labels


def ctc_greedy_decoder(probs_seq, blank=95, keep_blank_in_idxs=True):
    """
    CTC greedy (best path) decoder.
    """
    raw_str = np.argmax(np.array(probs_seq), axis=1)
    remove_blank_in_pos = None if keep_blank_in_idxs else blank
    dedup_str, keep_idx_list = get_keep_pos_idxs(
        raw_str, remove_blank=remove_blank_in_pos)
    dst_str = remove_blank(dedup_str, blank=blank)
    return dst_str, keep_idx_list


J
Jethong 已提交
91
def instance_ctc_greedy_decoder(gather_info, logits_map, pts_num=4):
J
Jethong 已提交
92 93
    _, _, C = logits_map.shape
    ys, xs = zip(*gather_info)
J
Jethong 已提交
94 95 96 97 98 99
    logits_seq = logits_map[list(ys), list(xs)]
    probs_seq = logits_seq
    labels = np.argmax(probs_seq, axis=1)
    dst_str = [k for k, v_ in groupby(labels) if k != C - 1]
    detal = len(gather_info) // (pts_num - 1)
    keep_idx_list = [0] + [detal * (i + 1) for i in range(pts_num - 2)] + [-1]
J
Jethong 已提交
100 101 102 103
    keep_gather_list = [gather_info[idx] for idx in keep_idx_list]
    return dst_str, keep_gather_list


J
Jethong 已提交
104 105 106 107
def ctc_decoder_for_image(gather_info_list,
                          logits_map,
                          Lexicon_Table,
                          pts_num=6):
J
Jethong 已提交
108 109 110
    """
    CTC decoder using multiple processes.
    """
J
Jethong 已提交
111 112
    decoder_str = []
    decoder_xys = []
J
Jethong 已提交
113
    for gather_info in gather_info_list:
J
Jethong 已提交
114 115 116 117 118 119 120 121 122 123
        if len(gather_info) < pts_num:
            continue
        dst_str, xys_list = instance_ctc_greedy_decoder(
            gather_info, logits_map, pts_num=pts_num)
        dst_str_readable = ''.join([Lexicon_Table[idx] for idx in dst_str])
        if len(dst_str_readable) < 2:
            continue
        decoder_str.append(dst_str_readable)
        decoder_xys.append(xys_list)
    return decoder_str, decoder_xys
J
Jethong 已提交
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


def sort_with_direction(pos_list, f_direction):
    """
    f_direction: h x w x 2
    pos_list: [[y, x], [y, x], [y, x] ...]
    """

    def sort_part_with_direction(pos_list, point_direction):
        pos_list = np.array(pos_list).reshape(-1, 2)
        point_direction = np.array(point_direction).reshape(-1, 2)
        average_direction = np.mean(point_direction, axis=0, keepdims=True)
        pos_proj_leng = np.sum(pos_list * average_direction, axis=1)
        sorted_list = pos_list[np.argsort(pos_proj_leng)].tolist()
        sorted_direction = point_direction[np.argsort(pos_proj_leng)].tolist()
        return sorted_list, sorted_direction

    pos_list = np.array(pos_list).reshape(-1, 2)
    point_direction = f_direction[pos_list[:, 0], pos_list[:, 1]]  # x, y
    point_direction = point_direction[:, ::-1]  # x, y -> y, x
    sorted_point, sorted_direction = sort_part_with_direction(pos_list,
                                                              point_direction)

    point_num = len(sorted_point)
    if point_num >= 16:
        middle_num = point_num // 2
        first_part_point = sorted_point[:middle_num]
        first_point_direction = sorted_direction[:middle_num]
        sorted_fist_part_point, sorted_fist_part_direction = sort_part_with_direction(
            first_part_point, first_point_direction)

        last_part_point = sorted_point[middle_num:]
        last_point_direction = sorted_direction[middle_num:]
        sorted_last_part_point, sorted_last_part_direction = sort_part_with_direction(
            last_part_point, last_point_direction)
        sorted_point = sorted_fist_part_point + sorted_last_part_point
        sorted_direction = sorted_fist_part_direction + sorted_last_part_direction

    return sorted_point, np.array(sorted_direction)


J
Jethong 已提交
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
def add_id(pos_list, image_id=0):
    """
    Add id for gather feature, for inference.
    """
    new_list = []
    for item in pos_list:
        new_list.append((image_id, item[0], item[1]))
    return new_list


def sort_and_expand_with_direction(pos_list, f_direction):
    """
    f_direction: h x w x 2
    pos_list: [[y, x], [y, x], [y, x] ...]
    """
    h, w, _ = f_direction.shape
    sorted_list, point_direction = sort_with_direction(pos_list, f_direction)

    point_num = len(sorted_list)
    sub_direction_len = max(point_num // 3, 2)
    left_direction = point_direction[:sub_direction_len, :]
    right_dirction = point_direction[point_num - sub_direction_len:, :]

    left_average_direction = -np.mean(left_direction, axis=0, keepdims=True)
    left_average_len = np.linalg.norm(left_average_direction)
    left_start = np.array(sorted_list[0])
    left_step = left_average_direction / (left_average_len + 1e-6)

    right_average_direction = np.mean(right_dirction, axis=0, keepdims=True)
    right_average_len = np.linalg.norm(right_average_direction)
    right_step = right_average_direction / (right_average_len + 1e-6)
    right_start = np.array(sorted_list[-1])

    append_num = max(
        int((left_average_len + right_average_len) / 2.0 * 0.15), 1)
    left_list = []
    right_list = []
    for i in range(append_num):
        ly, lx = np.round(left_start + left_step * (i + 1)).flatten().astype(
            'int32').tolist()
        if ly < h and lx < w and (ly, lx) not in left_list:
            left_list.append((ly, lx))
        ry, rx = np.round(right_start + right_step * (i + 1)).flatten().astype(
            'int32').tolist()
        if ry < h and rx < w and (ry, rx) not in right_list:
            right_list.append((ry, rx))

    all_list = left_list[::-1] + sorted_list + right_list
    return all_list


J
Jethong 已提交
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
def sort_and_expand_with_direction_v2(pos_list, f_direction, binary_tcl_map):
    """
    f_direction: h x w x 2
    pos_list: [[y, x], [y, x], [y, x] ...]
    binary_tcl_map: h x w
    """
    h, w, _ = f_direction.shape
    sorted_list, point_direction = sort_with_direction(pos_list, f_direction)

    point_num = len(sorted_list)
    sub_direction_len = max(point_num // 3, 2)
    left_direction = point_direction[:sub_direction_len, :]
    right_dirction = point_direction[point_num - sub_direction_len:, :]

    left_average_direction = -np.mean(left_direction, axis=0, keepdims=True)
    left_average_len = np.linalg.norm(left_average_direction)
    left_start = np.array(sorted_list[0])
    left_step = left_average_direction / (left_average_len + 1e-6)

    right_average_direction = np.mean(right_dirction, axis=0, keepdims=True)
    right_average_len = np.linalg.norm(right_average_direction)
    right_step = right_average_direction / (right_average_len + 1e-6)
    right_start = np.array(sorted_list[-1])

    append_num = max(
        int((left_average_len + right_average_len) / 2.0 * 0.15), 1)
    max_append_num = 2 * append_num

    left_list = []
    right_list = []
    for i in range(max_append_num):
        ly, lx = np.round(left_start + left_step * (i + 1)).flatten().astype(
            'int32').tolist()
        if ly < h and lx < w and (ly, lx) not in left_list:
            if binary_tcl_map[ly, lx] > 0.5:
                left_list.append((ly, lx))
            else:
                break

    for i in range(max_append_num):
        ry, rx = np.round(right_start + right_step * (i + 1)).flatten().astype(
            'int32').tolist()
        if ry < h and rx < w and (ry, rx) not in right_list:
            if binary_tcl_map[ry, rx] > 0.5:
                right_list.append((ry, rx))
            else:
                break

    all_list = left_list[::-1] + sorted_list + right_list
    return all_list


J
Jethong 已提交
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
def point_pair2poly(point_pair_list):
    """
    Transfer vertical point_pairs into poly point in clockwise.
    """
    point_num = len(point_pair_list) * 2
    point_list = [0] * point_num
    for idx, point_pair in enumerate(point_pair_list):
        point_list[idx] = point_pair[0]
        point_list[point_num - 1 - idx] = point_pair[1]
    return np.array(point_list).reshape(-1, 2)


def shrink_quad_along_width(quad, begin_width_ratio=0., end_width_ratio=1.):
    ratio_pair = np.array(
        [[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
    p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair
    p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair
    return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]])


def expand_poly_along_width(poly, shrink_ratio_of_width=0.3):
    """
    expand poly along width.
    """
    point_num = poly.shape[0]
    left_quad = np.array(
        [poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32)
    left_ratio = -shrink_ratio_of_width * np.linalg.norm(left_quad[0] - left_quad[3]) / \
                 (np.linalg.norm(left_quad[0] - left_quad[1]) + 1e-6)
    left_quad_expand = shrink_quad_along_width(left_quad, left_ratio, 1.0)
    right_quad = np.array(
        [
            poly[point_num // 2 - 2], poly[point_num // 2 - 1],
            poly[point_num // 2], poly[point_num // 2 + 1]
        ],
        dtype=np.float32)
    right_ratio = 1.0 + shrink_ratio_of_width * np.linalg.norm(right_quad[0] - right_quad[3]) / \
                  (np.linalg.norm(right_quad[0] - right_quad[1]) + 1e-6)
    right_quad_expand = shrink_quad_along_width(right_quad, 0.0, right_ratio)
    poly[0] = left_quad_expand[0]
    poly[-1] = left_quad_expand[-1]
    poly[point_num // 2 - 1] = right_quad_expand[1]
    poly[point_num // 2] = right_quad_expand[2]
    return poly


def restore_poly(instance_yxs_list, seq_strs, p_border, ratio_w, ratio_h, src_w,
                 src_h, valid_set):
    poly_list = []
    keep_str_list = []
    for yx_center_line, keep_str in zip(instance_yxs_list, seq_strs):
        if len(keep_str) < 2:
            print('--> too short, {}'.format(keep_str))
            continue

        offset_expand = 1.0
        if valid_set == 'totaltext':
            offset_expand = 1.2

        point_pair_list = []
        for y, x in yx_center_line:
            offset = p_border[:, y, x].reshape(2, 2) * offset_expand
            ori_yx = np.array([y, x], dtype=np.float32)
            point_pair = (ori_yx + offset)[:, ::-1] * 4.0 / np.array(
                [ratio_w, ratio_h]).reshape(-1, 2)
            point_pair_list.append(point_pair)

        detected_poly = point_pair2poly(point_pair_list)
        detected_poly = expand_poly_along_width(
            detected_poly, shrink_ratio_of_width=0.2)
        detected_poly[:, 0] = np.clip(detected_poly[:, 0], a_min=0, a_max=src_w)
        detected_poly[:, 1] = np.clip(detected_poly[:, 1], a_min=0, a_max=src_h)

        keep_str_list.append(keep_str)
        if valid_set == 'partvgg':
            middle_point = len(detected_poly) // 2
            detected_poly = detected_poly[
                [0, middle_point - 1, middle_point, -1], :]
            poly_list.append(detected_poly)
        elif valid_set == 'totaltext':
            poly_list.append(detected_poly)
J
Jethong 已提交
349
        else:
J
Jethong 已提交
350 351 352
            print('--> Not supported format.')
            exit(-1)
    return poly_list, keep_str_list
J
Jethong 已提交
353 354


J
Jethong 已提交
355 356 357 358 359
def generate_pivot_list_fast(p_score,
                             p_char_maps,
                             f_direction,
                             Lexicon_Table,
                             score_thresh=0.5):
J
Jethong 已提交
360 361 362 363 364
    """
    return center point and end point of TCL instance; filter with the char maps;
    """
    p_score = p_score[0]
    f_direction = f_direction.transpose(1, 2, 0)
J
Jethong 已提交
365 366 367
    ret, p_tcl_map = cv2.threshold(p_score, score_thresh, 255,
                                   cv2.THRESH_BINARY)
    skeleton_map = thin(p_tcl_map.astype('uint8'))
J
Jethong 已提交
368
    instance_count, instance_label_map = cv2.connectedComponents(
J
Jethong 已提交
369
        skeleton_map, connectivity=8)
J
Jethong 已提交
370 371 372 373 374 375 376 377 378

    # get TCL Instance
    all_pos_yxs = []
    if instance_count > 0:
        for instance_id in range(1, instance_count):
            pos_list = []
            ys, xs = np.where(instance_label_map == instance_id)
            pos_list = list(zip(ys, xs))

J
Jethong 已提交
379
            if len(pos_list) < 3:
J
Jethong 已提交
380 381
                continue

J
Jethong 已提交
382 383
            pos_list_sorted = sort_and_expand_with_direction_v2(
                pos_list, f_direction, p_tcl_map)
J
Jethong 已提交
384 385 386
            all_pos_yxs.append(pos_list_sorted)

    p_char_maps = p_char_maps.transpose([1, 2, 0])
J
Jethong 已提交
387 388 389 390 391
    decoded_str, keep_yxs_list = ctc_decoder_for_image(
        all_pos_yxs, logits_map=p_char_maps, Lexicon_Table=Lexicon_Table)
    return keep_yxs_list, decoded_str


J
Jethong 已提交
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458
def extract_main_direction(pos_list, f_direction):
    """
    f_direction: h x w x 2
    pos_list: [[y, x], [y, x], [y, x] ...]
    """
    pos_list = np.array(pos_list)
    point_direction = f_direction[pos_list[:, 0], pos_list[:, 1]]
    point_direction = point_direction[:, ::-1]  # x, y -> y, x
    average_direction = np.mean(point_direction, axis=0, keepdims=True)
    average_direction = average_direction / (
        np.linalg.norm(average_direction) + 1e-6)
    return average_direction


def sort_by_direction_with_image_id_deprecated(pos_list, f_direction):
    """
    f_direction: h x w x 2
    pos_list: [[id, y, x], [id, y, x], [id, y, x] ...]
    """
    pos_list_full = np.array(pos_list).reshape(-1, 3)
    pos_list = pos_list_full[:, 1:]
    point_direction = f_direction[pos_list[:, 0], pos_list[:, 1]]  # x, y
    point_direction = point_direction[:, ::-1]  # x, y -> y, x
    average_direction = np.mean(point_direction, axis=0, keepdims=True)
    pos_proj_leng = np.sum(pos_list * average_direction, axis=1)
    sorted_list = pos_list_full[np.argsort(pos_proj_leng)].tolist()
    return sorted_list


def sort_by_direction_with_image_id(pos_list, f_direction):
    """
    f_direction: h x w x 2
    pos_list: [[y, x], [y, x], [y, x] ...]
    """

    def sort_part_with_direction(pos_list_full, point_direction):
        pos_list_full = np.array(pos_list_full).reshape(-1, 3)
        pos_list = pos_list_full[:, 1:]
        point_direction = np.array(point_direction).reshape(-1, 2)
        average_direction = np.mean(point_direction, axis=0, keepdims=True)
        pos_proj_leng = np.sum(pos_list * average_direction, axis=1)
        sorted_list = pos_list_full[np.argsort(pos_proj_leng)].tolist()
        sorted_direction = point_direction[np.argsort(pos_proj_leng)].tolist()
        return sorted_list, sorted_direction

    pos_list = np.array(pos_list).reshape(-1, 3)
    point_direction = f_direction[pos_list[:, 1], pos_list[:, 2]]  # x, y
    point_direction = point_direction[:, ::-1]  # x, y -> y, x
    sorted_point, sorted_direction = sort_part_with_direction(pos_list,
                                                              point_direction)

    point_num = len(sorted_point)
    if point_num >= 16:
        middle_num = point_num // 2
        first_part_point = sorted_point[:middle_num]
        first_point_direction = sorted_direction[:middle_num]
        sorted_fist_part_point, sorted_fist_part_direction = sort_part_with_direction(
            first_part_point, first_point_direction)

        last_part_point = sorted_point[middle_num:]
        last_point_direction = sorted_direction[middle_num:]
        sorted_last_part_point, sorted_last_part_direction = sort_part_with_direction(
            last_part_point, last_point_direction)
        sorted_point = sorted_fist_part_point + sorted_last_part_point
        sorted_direction = sorted_fist_part_direction + sorted_last_part_direction

    return sorted_point