exam-segment-django/segment/sheet_resolve/analysis/sheet/choice_infer.py

# @Author  : lightXu
# @File    : choice_infer.py
import os
import traceback
import time
import random
from django.conf import settings
from segment.sheet_resolve.tools import utils, brain_api
from itertools import chain
import re
import numpy as np
import cv2
import xml.etree.cElementTree as ET
from segment.sheet_resolve.tools.utils import crop_region_direct, create_xml, infer_number, combine_char_in_raw_format
from sklearn.cluster import DBSCAN
from segment.sheet_resolve.analysis.sheet.ocr_sheet import ocr2sheet


def get_split_index(array, dif=0):
    array = np.array(array)
    interval_list = np.abs(array[1:] - array[:-1])
    split_index = [0]
    for i, interval in enumerate(interval_list):
        if dif:
            split_dif = dif
        else:
            split_dif = np.mean(interval_list)
        if interval > split_dif:
            split_index.append(i + 1)

    split_index.append(len(array))
    split_index = sorted(list(set(split_index)))
    return split_index


def adjust_choice_m(image, xe, ye):
    dilate = 1
    blur = 5

    # Convert to gray
    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

    if blur != 0:
        image = cv2.GaussianBlur(image, (blur, blur), 0)

    # Apply threshold to get image with only b&w (binarization)
    image = cv2.threshold(image, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]

    kernel = np.ones((ye, xe), np.uint8)  # y轴膨胀, x轴膨胀

    dst = cv2.dilate(image, kernel, iterations=1)

    (major, minor, _) = cv2.__version__.split(".")
    contours = cv2.findContours(dst, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    cnts = contours[0] if int(major) > 3 else contours[1]

    # _, cnts, hierarchy = cv2.findContours(dst, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    right_limit = 0
    bottom_limit = 0
    for cnt_id, cnt in enumerate(reversed(cnts)):
        x, y, w, h = cv2.boundingRect(cnt)
        if x + w > right_limit:
            right_limit = x + w

        if y + h > bottom_limit:
            bottom_limit = y + h

    return right_limit, bottom_limit


def find_digital(ocr_raw_list):
    pattern = r'\d+'
    x_list = []
    y_list = []
    digital_list = list()
    chars_list = list()
    height_list, width_list = list(), list()
    ocr_dict_list = combine_char_in_raw_format(ocr_raw_list)
    for i, ele in enumerate(ocr_dict_list):
        words = ele['words']
        words = words.replace(' ', '').upper()  # 去除空格

        digital_words_m = re.finditer(pattern, words)
        digital_index_list = [(m.group(), m.span()) for m in digital_words_m if m]
        chars_index = [ele for ele in range(0, len(ele['chars']))]
        digital_index_detail_list = []
        for letter_info in digital_index_list:
            number = letter_info[0]
            index_start = letter_info[1][0]
            index_end = letter_info[1][1] - 1
            char_start = ele['chars'][index_start]
            char_end = ele['chars'][index_end]

            if index_start == index_end:
                digital_index_detail_list += [index_start]
            else:
                digital_index_detail_list += chars_index[index_start:index_end + 1]

            letter_loc_xmin = int(char_start['location']['left'])
            letter_loc_ymin = min(int(char_start['location']['top']), int(char_end['location']['top']))
            letter_loc_xmax = int(char_end['location']['left']) + int(char_end['location']['width'])
            letter_loc_ymax = max(int(char_start['location']['top']) + int(char_start['location']['height']),
                                  int(char_end['location']['top']) + int(char_end['location']['height']))

            mid_x = letter_loc_xmin + (letter_loc_xmax - letter_loc_xmin) // 2
            mid_y = letter_loc_ymin + (letter_loc_ymax - letter_loc_ymin) // 2

            # print(number, (mid_x, mid_y))
            x_list.append(mid_x)
            y_list.append(mid_y)

            height_list.append(letter_loc_ymax - letter_loc_ymin)
            width_list.append(letter_loc_xmax - letter_loc_xmin)

            number_loc = (letter_loc_xmin, letter_loc_ymin, letter_loc_xmax, letter_loc_ymax, mid_x, mid_y)
            digital_list.append({"digital": int(number), "loc": number_loc})

        current_chars = [char for index, char in enumerate(ele['chars'])
                         if index not in digital_index_detail_list and char['char'] not in ['.', ',', '。', '、']]

        chars_list += current_chars
    d_mean_height = sum(height_list) // len(height_list)
    d_mean_width = sum(width_list) // len(width_list)

    # mean_height = max(height_list)
    # mean_width = max(width_list)
    # print(x_list)
    # print(y_list)
    return digital_list, chars_list, d_mean_height, d_mean_width


def cluster2choice_m_(cluster_list, m_h, m_w):
    numbers = [ele['digital'] for ele in cluster_list]

    loc_top_interval = (np.array([ele['loc'][3] for ele in cluster_list][1:]) -
                        np.array([ele['loc'][3] for ele in cluster_list][:-1]))

    split_index = [0]
    for i, interval in enumerate(loc_top_interval):
        if interval > m_h * 1.5:
            split_index.append(i + 1)

    split_index.append(len(cluster_list))
    split_index = sorted(list(set(split_index)))
    block_list = []
    for i in range(len(split_index) - 1):
        block = cluster_list[split_index[i]: split_index[i + 1]]

        xmin = min([ele["loc"][0] for ele in block])
        ymin = min([ele["loc"][1] for ele in block])
        xmax = max([ele["loc"][2] for ele in block])
        ymax = max([ele["loc"][3] for ele in block])

        numbers = [ele['digital'] for ele in block]

        choice_m = {"number": numbers, "loc": (xmin, ymin, xmax, ymax)}
        block_list.append(choice_m)

    return block_list


def cluster2choice_m(cluster_list, mean_width):
    # 比较x坐标，去掉误差值
    numbers_x = [ele['loc'][4] for ele in cluster_list]
    numbers_x_array = np.array(numbers_x)
    numbers_x_interval = np.abs((numbers_x_array[1:] - numbers_x_array[:-1]))
    error_index_superset = np.where(numbers_x_interval >= mean_width)[0]
    error_index_superset_interval = error_index_superset[1:] - error_index_superset[:-1]
    t_index = list(np.where(error_index_superset_interval > 1)[0] + 1)
    t_index.insert(0, 0)
    t_index.append(len(error_index_superset))
    error = []
    for i in range(0, len(t_index) - 1):
        a = t_index[i]
        b = t_index[i + 1]
        block = list(error_index_superset[a: b])
        error += block[1:]

    cluster_list = [ele for i, ele in enumerate(cluster_list) if i not in error]
    numbers = [ele['digital'] for ele in cluster_list]
    numbers_array = np.array(numbers)

    # numbers_y = [ele['loc'][5] for ele in cluster_list]
    # numbers_y_array = np.array(numbers_y)
    # numbers_y_interval = np.abs((numbers_y_array[1:] - numbers_y_array[:-1]))
    # split_index = [0]
    # for i, interval in enumerate(numbers_y_interval):
    #     if interval > np.mean(numbers_y_interval):
    #         split_index.append(i + 1)
    #
    # split_index.append(len(cluster_list))
    # split_index = sorted(list(set(split_index)))
    # for i in range(len(split_index) - 1):
    #     block = cluster_list[split_index[i]: split_index[i + 1]]
    #     block_numbers = numbers_array[split_index[i]: split_index[i + 1]]

    # 确定数字题号的位置，前提：同block题号是某等差数列的子集
    numbers_sum = numbers_array + np.flipud(numbers_array)

    counts = np.bincount(numbers_sum)
    mode_times = np.max(counts)
    mode_value = np.argmax(counts)

    if mode_times != len(numbers) and mode_times >= 2:
        # 启动题号补全

        number_interval_list = abs(numbers_array[1:] - numbers_array[:-1])
        number_interval_counts = np.bincount(number_interval_list)
        # number_interval_mode_times = np.max(number_interval_counts)
        number_interval_mode_value = np.argmax(number_interval_counts)

        suspect_index = np.where(numbers_sum != mode_value)[0]
        numbers_array_len = len(numbers_array)
        for suspect in suspect_index:
            if suspect == 0:
                cond_left = False
                cond_right = numbers_array[suspect + 1] == numbers_array[suspect] + number_interval_mode_value
            elif suspect == numbers_array_len - 1:
                cond_right = False
                cond_left = numbers_array[suspect - 1] == numbers_array[suspect] - number_interval_mode_value
            else:
                cond_left = numbers_array[suspect - 1] == numbers_array[suspect] - number_interval_mode_value
                cond_right = numbers_array[suspect + 1] == numbers_array[suspect] + number_interval_mode_value

            if cond_left or cond_right:
                pass
            else:
                numbers_array[suspect] = -1

        numbers_array = infer_number(numbers_array, number_interval_mode_value)  # 推断题号

    numbers_interval = np.abs(numbers_array[1:] - numbers_array[:-1])

    split_index = [0]
    for i, interval in enumerate(numbers_interval):
        if interval > np.mean(numbers_interval):
            split_index.append(i + 1)

    split_index.append(len(cluster_list))
    split_index = sorted(list(set(split_index)))
    block_list = []

    for i in range(len(split_index) - 1):
        block = cluster_list[split_index[i]: split_index[i + 1]]
        block_numbers = numbers_array[split_index[i]: split_index[i + 1]]

        xmin = min([ele["loc"][0] for ele in block])
        ymin = min([ele["loc"][1] for ele in block])
        xmax = max([ele["loc"][2] for ele in block])
        ymax = max([ele["loc"][3] for ele in block])
        mid_x = xmin + (xmax - xmin) // 2
        mid_y = ymin + (ymax - ymin) // 2

        choice_m = {"numbers": list(block_numbers), "loc": [xmin, ymin, xmax, ymax, mid_x, mid_y]}
        block_list.append(choice_m)

    return block_list


def cluster_and_anti_abnormal(image, xml_path, digital_list, chars_list,
                              mean_height, mean_width, choice_s_height, choice_s_width, limit_loc):
    limit_left, limit_top, limit_right, limit_bottom = limit_loc
    limit_width, limit_height = limit_right - limit_left, limit_bottom - limit_top
    arr = np.ones((len(digital_list), 2))
    for i, ele in enumerate(digital_list):
        arr[i] = np.array([ele["loc"][-2], ele["loc"][-1]])

    if choice_s_height != 0:
        eps = int(choice_s_height * 2)
    else:
        eps = int(mean_height * 2.5)
    print("eps: ", eps)
    db = DBSCAN(eps=eps, min_samples=2, metric='chebyshev').fit(arr)

    labels = db.labels_
    # print(labels)

    cluster_label = []
    for ele in labels:
        if ele not in cluster_label and ele != -1:
            cluster_label.append(ele)

    a_e_dict = {k: [] for k in cluster_label}
    choice_m_numbers_list = []
    for index, ele in enumerate(labels):
        if ele != -1:
            a_e_dict[ele].append(digital_list[index])

    for ele in cluster_label:
        cluster = a_e_dict[ele]
        choice_m_numbers_list += cluster2choice_m(cluster, mean_width)

    all_list_nums = [ele["numbers"] for ele in choice_m_numbers_list]
    all_nums_len = [len(ele) for ele in all_list_nums]
    all_nums = list(chain.from_iterable(all_list_nums))

    counts = np.bincount(np.array(all_nums_len))
    if np.max(counts) < 2:
        mode_value = max(all_nums_len)
    else:
        mode_value = np.argmax(counts)
        mode_value = all_nums_len[np.where(np.array(all_nums_len) == mode_value)[0][-1]]

    if mode_value > 1:  # 缺失补全
        error_index_list = list(np.where(np.array(all_nums_len) != mode_value)[0])

        all_height = [ele["loc"][3] - ele["loc"][1] for index, ele
                      in enumerate(choice_m_numbers_list) if index not in error_index_list]
        choice_m_mean_height = int(sum(all_height) / len(all_height))

        for e_index in list(error_index_list):
            current_choice_m = choice_m_numbers_list[e_index]
            current_numbers_list = list(all_list_nums[e_index])
            current_len = all_nums_len[e_index]
            dif = mode_value - current_len

            if 1 in current_numbers_list:
                t2 = current_numbers_list + [-1] * dif
                infer_t1_list = infer_number(t2)  # 后补
                infer_t2_list = infer_number(t2)  # 后补
                cond1 = False
                cond2 = True
            else:
                t1_cond = [True] * dif
                t2_cond = [True] * dif

                t1 = [-1] * dif + current_numbers_list
                infer_t1_list = infer_number(t1)  # 前补
                t2 = current_numbers_list + [-1] * dif
                infer_t2_list = infer_number(t2)  # 后补

                for i in range(0, dif):
                    t1_infer = infer_t1_list[i]
                    t2_infer = infer_t2_list[-i - 1]
                    if t1_infer == 0 or t1_infer in all_nums:
                        t1_cond[i] = False
                    if t2_infer in all_nums:
                        t2_cond[i] = False
                cond1 = not (False in t1_cond)
                cond2 = not (False in t2_cond)

            if cond1 and not cond2:
                current_loc = current_choice_m["loc"]
                current_height = current_loc[3] - current_loc[1]

                infer_height = max((choice_m_mean_height - current_height), int(dif * current_height / current_len))
                choice_m_numbers_list[e_index]["loc"][1] = current_loc[1] - infer_height
                choice_m_numbers_list[e_index]["loc"][5] = (choice_m_numbers_list[e_index]["loc"][1] +
                                                            (choice_m_numbers_list[e_index]["loc"][3] -
                                                             choice_m_numbers_list[e_index]["loc"][1]) // 2)
                choice_m_numbers_list[e_index]["numbers"] = infer_t1_list
                all_nums.extend(infer_t1_list[:dif])
            if not cond1 and cond2:
                current_loc = current_choice_m["loc"]
                current_height = current_loc[3] - current_loc[1]

                infer_height = max((choice_m_mean_height - current_height), int(dif * current_height / current_len))
                infer_bottom = min(current_loc[3] + infer_height, limit_height-1)
                if infer_bottom <= limit_height:
                    choice_m_numbers_list[e_index]["loc"][3] = infer_bottom
                    choice_m_numbers_list[e_index]["loc"][5] = (choice_m_numbers_list[e_index]["loc"][1] +
                                                                (choice_m_numbers_list[e_index]["loc"][3] -
                                                                 choice_m_numbers_list[e_index]["loc"][1]) // 2)
                    choice_m_numbers_list[e_index]["numbers"] = infer_t2_list
                    all_nums.extend(infer_t2_list[-dif:])
            else:
                # cond1 = cond2 = true, 因为infer选择题时已横向排序， 默认这种情况不会出现
                pass

    for ele in choice_m_numbers_list:
        loc = ele["loc"]
        if loc[3] - loc[1] >= loc[2] - loc[0]:
            direction = 180
        else:
            direction = 90
        ele.update({'direction': direction})
    # tree = ET.parse(xml_path)
    # for index, choice_m in enumerate(choice_m_numbers_list):
    #     name = str(choice_m["numbers"])
    #     xmin, ymin, xmax, ymax, _, _ = choice_m["loc"]
    #     tree = create_xml(name, tree, str(xmin + limit_left), str(ymin + limit_top), str(xmax + limit_left), str(ymax + limit_top))
    #
    # tree.write(xml_path)

    choice_m_numbers_list = sorted(choice_m_numbers_list, key=lambda x: x['loc'][3] - x['loc'][1], reverse=True)
    choice_m_numbers_right_limit = max([ele['loc'][2] for ele in choice_m_numbers_list])
    remain_len = len(choice_m_numbers_list)
    choice_m_list = list()
    need_revised_choice_m_list = list()
    while remain_len > 0:
        # 先确定属于同行的数据，然后找字母划分block
        # random_index = random.randint(0, len(choice_m_numbers_list)-1)
        random_index = 0
        # print(random_index)
        ymax_limit = choice_m_numbers_list[random_index]["loc"][3]
        ymin_limit = choice_m_numbers_list[random_index]["loc"][1]
        # choice_m_numbers_list.pop(random_index)

        # 当前行的choice_m
        current_row_choice_m_d = [ele for ele in choice_m_numbers_list if ymin_limit < ele["loc"][5] < ymax_limit]
        current_row_choice_m_d = sorted(current_row_choice_m_d, key=lambda x: x["loc"][0])
        # current_row_choice_m_d.append(choice_m_numbers_list[random_index])
        split_pix = sorted([ele["loc"][0] for ele in current_row_choice_m_d])  # xmin排序
        split_index = get_split_index(split_pix)
        split_pix = [split_pix[ele] for ele in split_index[:-1]]

        block_list = []
        for i in range(len(split_index) - 1):
            block = current_row_choice_m_d[split_index[i]: split_index[i + 1]]
            if len(block) > 1:
                remain_len = remain_len - (len(block) - 1)
                numbers_new = []
                loc_new = [[], [], [], []]
                for blk in block:
                    loc_old = blk["loc"]
                    numbers_new.extend(blk["numbers"])
                    for ii in range(4):
                        loc_new[ii].append(loc_old[ii])

                loc_new[0] = min(loc_new[0])
                loc_new[1] = min(loc_new[1])
                loc_new[2] = max(loc_new[2])
                loc_new[3] = max(loc_new[3])

                loc_new.append(loc_new[0] + (loc_new[2] - loc_new[0]) // 2)
                loc_new.append(loc_new[1] + (loc_new[3] - loc_new[1]) // 2)

                block = [{"numbers": sorted(numbers_new), "loc": loc_new, "direction": block[0]["direction"]}]

            block_list.extend(block)

        current_row_choice_m_d = block_list
        current_row_chars = [ele for ele in chars_list
                             if ymin_limit < (ele["location"]["top"] + ele["location"]["height"] // 2) < ymax_limit]

        # if not current_row_chars:
        #     max_char_width = choice_s_width // 4
        #     row_chars_xmax = choice_m_numbers_right_limit + int(choice_s_width * 1.5)
        # else:
        #     max_char_width = max([ele["location"]["width"] for ele in current_row_chars]) // 2
        #     row_chars_xmax = max(
        #         [ele["location"]["left"] + ele["location"]["width"] for ele in current_row_chars]) + max_char_width * 2

        # split_index.append(row_chars_xmax)  # 边界
        split_pix.append(round(split_pix[-1] + choice_s_width * 1.2))
        for i in range(0, len(split_index) - 1):
            left_limit = split_index[i]
            right_limit = split_index[i + 1]
            block_chars = [ele for ele in current_row_chars
                           if left_limit < (ele["location"]["left"] + ele["location"]["width"] // 2) < right_limit]

            # chars_xmin = min([ele["location"]["left"] for ele in block_chars]) - max_char_width
            # chars_xmax = max(
            #     [ele["location"]["left"] + ele["location"]["width"] for ele in block_chars]) + max_char_width

            # a_z = '_ABCD_FGH__K_MNOPQRSTUVWXYZ'  EIJL -> _
            # a_z = '_ABCDEFGHI_K_MNOPQRSTUVWXYZ'
            a_z = '_ABCD_FGHT'
            # letter_text = set([ele['char'].upper() for ele in block_chars if ele['char'].upper() in a_z])
            letter_index = [a_z.index(ele['char'].upper()) for ele in block_chars if ele['char'].upper() in a_z]

            letter_index_times = {ele: 0 for ele in set(letter_index)}
            for l_index in letter_index:
                letter_index_times[l_index] += 1

            if (a_z.index("T") in letter_index) and (a_z.index("F") in letter_index):
                choice_option = "T, F"
                cols = 2
            else:
                if len(letter_index) < 1:
                    tmp = 4
                    choice_option = 'A,B,C,D'
                else:
                    tmp = max(set(letter_index))
                # while letter_index_times[tmp] < 2 and tmp > 3:
                #     t_list = list(set(letter_index))
                #     t_list.remove(tmp)
                #     tmp = max(t_list)

                    choice_option = ",".join(a_z[min(letter_index):tmp + 1])
                cols = tmp

            bias = 3  # pix
            current_loc = current_row_choice_m_d[i]["loc"]
            location = dict(xmin=(current_loc[2] + bias) + limit_left,  # 当前数字xmax右边
                            # xmin=max(current_loc[2] + bias, chars_xmin) + limit_left,
                            ymin=current_loc[1] + limit_top,

                            xmax=(right_limit - bias) + limit_left,
                            # xmax=min(chars_xmax, right_limit - bias) + limit_left,
                            ymax=current_loc[3] + limit_top)

            try:
                choice_m_img = utils.crop_region(image, location)
                right_loc, bottom_loc = adjust_choice_m(choice_m_img, mean_height, mean_width * 2)
                if right_loc > 0:
                    location.update(dict(xmax=right_loc + location['xmin']))
                if bottom_loc > 0:
                    location.update(dict(ymax=bottom_loc + location['ymin']))
            except Exception as e:
                print(e)
                traceback.print_exc()

            tmp_w, tmp_h = location['xmax'] - location['xmin'], location['ymax'] - location['ymin'],
            numbers = current_row_choice_m_d[i]["numbers"]
            direction = current_row_choice_m_d[i]["direction"]
            if direction == 180:
                choice_m = dict(class_name='choice_m',
                                number=numbers,
                                bounding_box=location,
                                choice_option=choice_option,
                                default_points=[5] * len(numbers),
                                direction=direction,
                                cols=cols,
                                rows=len(numbers),
                                single_width=tmp_w // cols,
                                single_height=tmp_h // len(numbers))
            else:
                choice_m = dict(class_name='choice_m',
                                number=numbers,
                                bounding_box=location,
                                choice_option=choice_option,
                                default_points=[5] * len(numbers),
                                direction=direction,
                                cols=len(numbers),
                                rows=cols,
                                single_width=tmp_w // len(numbers),
                                single_height=tmp_h // cols
                                )

            if tmp_w > 2 * choice_s_width:
                need_revised_choice_m_list.append(choice_m)
            else:
                choice_m_list.append(choice_m)

        remain_len = remain_len - len(current_row_choice_m_d)
        for ele in choice_m_numbers_list.copy():
            if ele in current_row_choice_m_d:
                choice_m_numbers_list.remove(ele)

        for ele in choice_m_numbers_list.copy():
            if ele in current_row_chars:
                choice_m_numbers_list.remove(ele)

    # 单独一行不聚类
    for i, revised_choice_m in enumerate(need_revised_choice_m_list):
        loc = revised_choice_m['bounding_box']
        left_part_loc = loc.copy()
        left_part_loc.update({'xmax': loc['xmin']+choice_s_width})
        choice_m_img = utils.crop_region(image, left_part_loc)
        right_loc, bottom_loc = adjust_choice_m(choice_m_img, mean_height, mean_width * 2)
        if right_loc > 0:
            left_part_loc.update(dict(xmax=right_loc + left_part_loc['xmin']))
        if bottom_loc > 0:
            left_part_loc.update(dict(ymax=bottom_loc + left_part_loc['ymin']))

        left_tmp_height = left_part_loc['ymax'] - left_part_loc['ymin']

        right_part_loc = loc.copy()
        # right_part_loc.update({'xmin': loc['xmax']-choice_s_width})
        right_part_loc.update({'xmin': left_part_loc['xmax']+5})
        choice_m_img = utils.crop_region(image, right_part_loc)
        right_loc, bottom_loc = adjust_choice_m(choice_m_img, mean_height, mean_width * 2)
        if right_loc > 0:
            right_part_loc.update(dict(xmax=right_loc + right_part_loc['xmin']))
        if bottom_loc > 0:
            right_part_loc.update(dict(ymax=bottom_loc + right_part_loc['ymin']))

        right_tmp_height = right_part_loc['ymax'] - right_part_loc['ymin']

        number_len = max(1, int(revised_choice_m['rows'] // (left_tmp_height // right_tmp_height)))
        number = [ele+revised_choice_m['number'][-1]+1 for ele in range(number_len)]
        rows = len(number)

        revised_choice_m.update({'bounding_box': left_part_loc})
        choice_m_list.append(revised_choice_m)

        tmp = revised_choice_m.copy()
        tmp.update({'bounding_box': right_part_loc, 'number': number, 'rows': rows})
        choice_m_list.append(tmp)

    tmp = choice_m_list.copy()
    for ele in tmp:
        loc = ele["bounding_box"]
        w, h = loc['xmax'] - loc['xmin'], loc['ymax'] - loc['ymin']
        if w*h < choice_s_width*choice_s_height:
            choice_m_list.remove(ele)
    return choice_m_list


def infer_choice_m(image, tf_sheet, ocr, xml=None):
    infer_box_list = ocr2sheet(image, tf_sheet, ocr, xml)
    # print(sheet_region_list)
    choice_m_list = []

    choice_s_h_list = [int(ele['bounding_box']['ymax']) - int(ele['bounding_box']['ymin']) for ele in tf_sheet
                       if ele['class_name'] == 'choice_s']
    if choice_s_h_list:
        choice_s_height = sum(choice_s_h_list) // len(choice_s_h_list)
    else:
        choice_s_height = 0

    choice_s_w_list = [int(ele['bounding_box']['xmax']) - int(ele['bounding_box']['xmin']) for ele in tf_sheet
                       if ele['class_name'] == 'choice_s']
    if choice_s_w_list:
        choice_s_width = sum(choice_s_w_list) // len(choice_s_w_list)

    else:
        choice_s_width = 0

    for infer_box in infer_box_list:
        # {'loc': [240, 786, 1569, 1368]}
        loc = infer_box['loc']
        xmin, ymin, xmax, ymax = loc[0], loc[1], loc[2], loc[3]
        choice_flag = False

        for ele in tf_sheet:
            if ele['class_name'] in ['choice_m', 'choice_s']:
                tf_loc = ele['bounding_box']
                tf_loc_l = tf_loc['xmin']
                tf_loc_t = tf_loc['ymin']
                if xmin < tf_loc_l < xmax and ymin < tf_loc_t < ymax:
                    choice_flag = True
                    break

        if choice_flag:
            infer_image = utils.crop_region_direct(image, loc)
            try:
                save_dir = os.path.join(settings.MEDIA_ROOT, 'tmp')
                if not os.path.exists(save_dir):
                    os.makedirs(save_dir)
                save_path = os.path.join(save_dir, 'choice.jpeg')
                cv2.imwrite(save_path, infer_image)
                img_tmp = utils.read_single_img(save_path)
                os.remove(save_path)
                ocr = brain_api.get_ocr_text_and_coordinate(img_tmp, 'accurate', 'CHN_ENG')
            except Exception as e:
                print('write choice and ocr failed')
                traceback.print_exc()
                ocr = brain_api.get_ocr_text_and_coordinate(infer_image, 'accurate', 'CHN_ENG')

            try:
                digital_list, chars_list, digital_mean_h, digital_mean_w = find_digital(ocr)
                choice_m = cluster_and_anti_abnormal(image, xml, digital_list, chars_list,
                                                     digital_mean_h, digital_mean_w,
                                                     choice_s_height, choice_s_width, loc)

                choice_m_list.extend(choice_m)
            except Exception as e:
                traceback.print_exc()
                print('not found choice feature')
                pass

    # print(choice_m_list)
    # tf_choice_sheet = [ele for ele in tf_sheet if ele['class_name'] == 'choice_m']

    sheet_tmp = choice_m_list.copy()
    remove_index = []
    for i, region in enumerate(sheet_tmp):
        if i not in remove_index:
            box = region['bounding_box']
            for j, region_in in enumerate(sheet_tmp):
                box_in = region_in['bounding_box']
                iou = utils.cal_iou(box, box_in)
                if iou[0] > 0.85 and i != j:
                    choice_m_list.remove(region)
                    remove_index.append(j)
                    break

    return choice_m_list