sast_postprocess.py 13.3 KB
Newer Older
M
MissPenguin 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os
import sys
J
Jethong 已提交
21

M
MissPenguin 已提交
22 23 24 25 26 27
__dir__ = os.path.dirname(__file__)
sys.path.append(__dir__)
sys.path.append(os.path.join(__dir__, '..'))

import numpy as np
from .locality_aware_nms import nms_locality
M
MissPenguin 已提交
28
import paddle
M
MissPenguin 已提交
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
import cv2
import time


class SASTPostProcess(object):
    """
    The post process for SAST.
    """

    def __init__(self,
                 score_thresh=0.5,
                 nms_thresh=0.2,
                 sample_pts_num=2,
                 shrink_ratio_of_width=0.3,
                 expand_scale=1.0,
                 tcl_map_thresh=0.5,
                 **kwargs):

        self.score_thresh = score_thresh
        self.nms_thresh = nms_thresh
        self.sample_pts_num = sample_pts_num
        self.shrink_ratio_of_width = shrink_ratio_of_width
        self.expand_scale = expand_scale
        self.tcl_map_thresh = tcl_map_thresh
J
Jethong 已提交
53

M
MissPenguin 已提交
54 55 56 57
        # c++ la-nms is faster, but only support python 3.5
        self.is_python35 = False
        if sys.version_info.major == 3 and sys.version_info.minor == 5:
            self.is_python35 = True
J
Jethong 已提交
58

M
MissPenguin 已提交
59 60 61 62 63 64 65 66 67 68 69
    def point_pair2poly(self, point_pair_list):
        """
        Transfer vertical point_pairs into poly point in clockwise.
        """
        # constract poly
        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)
J
Jethong 已提交
70 71 72 73 74

    def shrink_quad_along_width(self,
                                quad,
                                begin_width_ratio=0.,
                                end_width_ratio=1.):
M
MissPenguin 已提交
75 76 77
        """ 
        Generate shrink_quad_along_width.
        """
J
Jethong 已提交
78 79
        ratio_pair = np.array(
            [[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
M
MissPenguin 已提交
80 81 82
        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]])
J
Jethong 已提交
83

M
MissPenguin 已提交
84 85 86 87 88
    def expand_poly_along_width(self, poly, shrink_ratio_of_width=0.3):
        """
        expand poly along width.
        """
        point_num = poly.shape[0]
J
Jethong 已提交
89 90
        left_quad = np.array(
            [poly[0], poly[1], poly[-2], poly[-1]], dtype=np.float32)
M
MissPenguin 已提交
91
        left_ratio = -shrink_ratio_of_width * np.linalg.norm(left_quad[0] - left_quad[3]) / \
J
Jethong 已提交
92 93 94 95 96 97 98 99 100
                     (np.linalg.norm(left_quad[0] - left_quad[1]) + 1e-6)
        left_quad_expand = self.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)
M
MissPenguin 已提交
101
        right_ratio = 1.0 + \
J
Jethong 已提交
102 103 104 105
                      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 = self.shrink_quad_along_width(right_quad, 0.0,
                                                         right_ratio)
M
MissPenguin 已提交
106 107 108 109 110 111 112 113 114
        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_quad(self, tcl_map, tcl_map_thresh, tvo_map):
        """Restore quad."""
        xy_text = np.argwhere(tcl_map[:, :, 0] > tcl_map_thresh)
J
Jethong 已提交
115
        xy_text = xy_text[:, ::-1]  # (n, 2)
M
MissPenguin 已提交
116 117 118 119 120 121 122 123 124 125 126

        # Sort the text boxes via the y axis
        xy_text = xy_text[np.argsort(xy_text[:, 1])]

        scores = tcl_map[xy_text[:, 1], xy_text[:, 0], 0]
        scores = scores[:, np.newaxis]

        # Restore
        point_num = int(tvo_map.shape[-1] / 2)
        assert point_num == 4
        tvo_map = tvo_map[xy_text[:, 1], xy_text[:, 0], :]
J
Jethong 已提交
127
        xy_text_tile = np.tile(xy_text, (1, point_num))  # (n, point_num * 2)
M
MissPenguin 已提交
128 129 130 131 132 133 134 135
        quads = xy_text_tile - tvo_map

        return scores, quads, xy_text

    def quad_area(self, quad):
        """
        compute area of a quad.
        """
J
Jethong 已提交
136 137 138 139
        edge = [(quad[1][0] - quad[0][0]) * (quad[1][1] + quad[0][1]),
                (quad[2][0] - quad[1][0]) * (quad[2][1] + quad[1][1]),
                (quad[3][0] - quad[2][0]) * (quad[3][1] + quad[2][1]),
                (quad[0][0] - quad[3][0]) * (quad[0][1] + quad[3][1])]
M
MissPenguin 已提交
140
        return np.sum(edge) / 2.
J
Jethong 已提交
141

M
MissPenguin 已提交
142 143 144 145 146 147 148 149 150 151 152 153
    def nms(self, dets):
        if self.is_python35:
            import lanms
            dets = lanms.merge_quadrangle_n9(dets, self.nms_thresh)
        else:
            dets = nms_locality(dets, self.nms_thresh)
        return dets

    def cluster_by_quads_tco(self, tcl_map, tcl_map_thresh, quads, tco_map):
        """
        Cluster pixels in tcl_map based on quads.
        """
J
Jethong 已提交
154
        instance_count = quads.shape[0] + 1  # contain background
M
MissPenguin 已提交
155 156 157 158 159 160 161
        instance_label_map = np.zeros(tcl_map.shape[:2], dtype=np.int32)
        if instance_count == 1:
            return instance_count, instance_label_map

        # predict text center
        xy_text = np.argwhere(tcl_map[:, :, 0] > tcl_map_thresh)
        n = xy_text.shape[0]
J
Jethong 已提交
162 163
        xy_text = xy_text[:, ::-1]  # (n, 2)
        tco = tco_map[xy_text[:, 1], xy_text[:, 0], :]  # (n, 2)
M
MissPenguin 已提交
164
        pred_tc = xy_text - tco
J
Jethong 已提交
165

M
MissPenguin 已提交
166 167
        # get gt text center
        m = quads.shape[0]
J
Jethong 已提交
168
        gt_tc = np.mean(quads, axis=1)  # (m, 2)
M
MissPenguin 已提交
169

J
Jethong 已提交
170 171 172 173 174
        pred_tc_tile = np.tile(pred_tc[:, np.newaxis, :],
                               (1, m, 1))  # (n, m, 2)
        gt_tc_tile = np.tile(gt_tc[np.newaxis, :, :], (n, 1, 1))  # (n, m, 2)
        dist_mat = np.linalg.norm(pred_tc_tile - gt_tc_tile, axis=2)  # (n, m)
        xy_text_assign = np.argmin(dist_mat, axis=1) + 1  # (n,)
M
MissPenguin 已提交
175 176 177 178 179 180 181 182

        instance_label_map[xy_text[:, 1], xy_text[:, 0]] = xy_text_assign
        return instance_count, instance_label_map

    def estimate_sample_pts_num(self, quad, xy_text):
        """
        Estimate sample points number.
        """
J
Jethong 已提交
183 184 185 186
        eh = (np.linalg.norm(quad[0] - quad[3]) +
              np.linalg.norm(quad[1] - quad[2])) / 2.0
        ew = (np.linalg.norm(quad[0] - quad[1]) +
              np.linalg.norm(quad[2] - quad[3])) / 2.0
M
MissPenguin 已提交
187 188

        dense_sample_pts_num = max(2, int(ew))
J
Jethong 已提交
189 190 191 192 193 194 195 196 197 198 199 200
        dense_xy_center_line = xy_text[np.linspace(
            0,
            xy_text.shape[0] - 1,
            dense_sample_pts_num,
            endpoint=True,
            dtype=np.float32).astype(np.int32)]

        dense_xy_center_line_diff = dense_xy_center_line[
            1:] - dense_xy_center_line[:-1]
        estimate_arc_len = np.sum(
            np.linalg.norm(
                dense_xy_center_line_diff, axis=1))
M
MissPenguin 已提交
201 202 203 204

        sample_pts_num = max(2, int(estimate_arc_len / eh))
        return sample_pts_num

J
Jethong 已提交
205 206 207 208 209 210 211 212 213 214 215 216 217
    def detect_sast(self,
                    tcl_map,
                    tvo_map,
                    tbo_map,
                    tco_map,
                    ratio_w,
                    ratio_h,
                    src_w,
                    src_h,
                    shrink_ratio_of_width=0.3,
                    tcl_map_thresh=0.5,
                    offset_expand=1.0,
                    out_strid=4.0):
M
MissPenguin 已提交
218 219 220 221
        """
        first resize the tcl_map, tvo_map and tbo_map to the input_size, then restore the polys
        """
        # restore quad
J
Jethong 已提交
222 223
        scores, quads, xy_text = self.restore_quad(tcl_map, tcl_map_thresh,
                                                   tvo_map)
M
MissPenguin 已提交
224 225 226 227 228 229 230 231 232 233 234 235 236
        dets = np.hstack((quads, scores)).astype(np.float32, copy=False)
        dets = self.nms(dets)
        if dets.shape[0] == 0:
            return []
        quads = dets[:, :-1].reshape(-1, 4, 2)

        # Compute quad area
        quad_areas = []
        for quad in quads:
            quad_areas.append(-self.quad_area(quad))

        # instance segmentation
        # instance_count, instance_label_map = cv2.connectedComponents(tcl_map.astype(np.uint8), connectivity=8)
J
Jethong 已提交
237 238
        instance_count, instance_label_map = self.cluster_by_quads_tco(
            tcl_map, tcl_map_thresh, quads, tco_map)
M
MissPenguin 已提交
239 240 241 242 243 244 245 246 247

        # restore single poly with tcl instance.
        poly_list = []
        for instance_idx in range(1, instance_count):
            xy_text = np.argwhere(instance_label_map == instance_idx)[:, ::-1]
            quad = quads[instance_idx - 1]
            q_area = quad_areas[instance_idx - 1]
            if q_area < 5:
                continue
J
Jethong 已提交
248

M
MissPenguin 已提交
249
            #
J
Jethong 已提交
250 251
            len1 = float(np.linalg.norm(quad[0] - quad[1]))
            len2 = float(np.linalg.norm(quad[1] - quad[2]))
M
MissPenguin 已提交
252 253 254 255 256 257 258 259 260
            min_len = min(len1, len2)
            if min_len < 3:
                continue

            # filter small CC
            if xy_text.shape[0] <= 0:
                continue

            # filter low confidence instance
J
Jethong 已提交
261
            xy_text_scores = tcl_map[xy_text[:, 1], xy_text[:, 0], 0]
M
MissPenguin 已提交
262
            if np.sum(xy_text_scores) / quad_areas[instance_idx - 1] < 0.1:
J
Jethong 已提交
263
                # if np.sum(xy_text_scores) / quad_areas[instance_idx - 1] < 0.05:
M
MissPenguin 已提交
264 265 266
                continue

            # sort xy_text
J
Jethong 已提交
267 268 269 270 271 272
            left_center_pt = np.array(
                [[(quad[0, 0] + quad[-1, 0]) / 2.0,
                  (quad[0, 1] + quad[-1, 1]) / 2.0]])  # (1, 2)
            right_center_pt = np.array(
                [[(quad[1, 0] + quad[2, 0]) / 2.0,
                  (quad[1, 1] + quad[2, 1]) / 2.0]])  # (1, 2)
M
MissPenguin 已提交
273 274 275 276 277 278 279 280 281 282
            proj_unit_vec = (right_center_pt - left_center_pt) / \
                            (np.linalg.norm(right_center_pt - left_center_pt) + 1e-6)
            proj_value = np.sum(xy_text * proj_unit_vec, axis=1)
            xy_text = xy_text[np.argsort(proj_value)]

            # Sample pts in tcl map
            if self.sample_pts_num == 0:
                sample_pts_num = self.estimate_sample_pts_num(quad, xy_text)
            else:
                sample_pts_num = self.sample_pts_num
J
Jethong 已提交
283 284 285 286 287 288
            xy_center_line = xy_text[np.linspace(
                0,
                xy_text.shape[0] - 1,
                sample_pts_num,
                endpoint=True,
                dtype=np.float32).astype(np.int32)]
M
MissPenguin 已提交
289 290 291 292 293 294

            point_pair_list = []
            for x, y in xy_center_line:
                # get corresponding offset
                offset = tbo_map[y, x, :].reshape(2, 2)
                if offset_expand != 1.0:
J
Jethong 已提交
295 296 297 298 299 300
                    offset_length = np.linalg.norm(
                        offset, axis=1, keepdims=True)
                    expand_length = np.clip(
                        offset_length * (offset_expand - 1),
                        a_min=0.5,
                        a_max=3.0)
M
MissPenguin 已提交
301
                    offset_detal = offset / offset_length * expand_length
J
Jethong 已提交
302 303
                    offset = offset + offset_detal
                    # original point
M
MissPenguin 已提交
304
                ori_yx = np.array([y, x], dtype=np.float32)
J
Jethong 已提交
305 306
                point_pair = (ori_yx + offset)[:, ::-1] * out_strid / np.array(
                    [ratio_w, ratio_h]).reshape(-1, 2)
M
MissPenguin 已提交
307 308 309 310
                point_pair_list.append(point_pair)

            # ndarry: (x, 2), expand poly along width
            detected_poly = self.point_pair2poly(point_pair_list)
J
Jethong 已提交
311 312 313 314 315 316
            detected_poly = self.expand_poly_along_width(detected_poly,
                                                         shrink_ratio_of_width)
            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)
M
MissPenguin 已提交
317 318 319 320
            poly_list.append(detected_poly)

        return poly_list

J
Jethong 已提交
321
    def __call__(self, outs_dict, shape_list):
M
MissPenguin 已提交
322 323 324 325
        score_list = outs_dict['f_score']
        border_list = outs_dict['f_border']
        tvo_list = outs_dict['f_tvo']
        tco_list = outs_dict['f_tco']
M
MissPenguin 已提交
326 327 328 329 330
        if isinstance(score_list, paddle.Tensor):
            score_list = score_list.numpy()
            border_list = border_list.numpy()
            tvo_list = tvo_list.numpy()
            tco_list = tco_list.numpy()
J
Jethong 已提交
331

M
MissPenguin 已提交
332 333 334
        img_num = len(shape_list)
        poly_lists = []
        for ino in range(img_num):
J
Jethong 已提交
335 336 337 338
            p_score = score_list[ino].transpose((1, 2, 0))
            p_border = border_list[ino].transpose((1, 2, 0))
            p_tvo = tvo_list[ino].transpose((1, 2, 0))
            p_tco = tco_list[ino].transpose((1, 2, 0))
M
MissPenguin 已提交
339 340
            src_h, src_w, ratio_h, ratio_w = shape_list[ino]

J
Jethong 已提交
341 342 343 344 345 346 347 348 349 350 351 352
            poly_list = self.detect_sast(
                p_score,
                p_tvo,
                p_border,
                p_tco,
                ratio_w,
                ratio_h,
                src_w,
                src_h,
                shrink_ratio_of_width=self.shrink_ratio_of_width,
                tcl_map_thresh=self.tcl_map_thresh,
                offset_expand=self.expand_scale)
M
MissPenguin 已提交
353 354 355
            poly_lists.append({'points': np.array(poly_list)})

        return poly_lists