FORMAT_CONVERSION_MAXCOMPUTE/format_convert/convert_image.py

# encoding=utf8
import inspect
import io
import logging
import os
import sys
import time
import requests
import numpy as np
from ocr.paddleocr import PaddleOCR

sys.path.append(os.path.dirname(__file__) + "/../")
from pdfminer.layout import LTLine
import traceback
import cv2
from format_convert import get_memory_info, _global
from format_convert.utils import judge_error_code, add_div, LineTable, get_table_html, get_logger, log, memory_decorator
from format_convert.table_correct import get_rotated_image
from format_convert.convert_need_interface import from_otr_interface, from_ocr_interface, from_gpu_interface_redis
from otr.table_line import table_preprocess, table_postprocess


def image_process(image_np, image_path, is_from_pdf=False, is_from_docx=False, use_ocr=True):
    from format_convert.convert_tree import _Table, _Sentence

    def get_cluster(t_list, b_list, axis):
        zip_list = list(zip(t_list, b_list))
        if len(zip_list) == 0:
            return t_list, b_list
        if len(zip_list[0]) > 0:
            zip_list.sort(key=lambda x: x[1][axis][1])
        cluster_list = []
        margin = 5
        for text, bbox in zip_list:
            _find = 0
            for cluster in cluster_list:
                if abs(cluster[1] - bbox[axis][1]) <= margin:
                    cluster[0].append([text, bbox])
                    cluster[1] = bbox[axis][1]
                    _find = 1
                    break
            if not _find:
                cluster_list.append([[[text, bbox]], bbox[axis][1]])

        new_text_list = []
        new_bbox_list = []
        for cluster in cluster_list:
            # print("=============convert_image")
            # print("cluster_list", cluster)
            center_y = 0
            for text, bbox in cluster[0]:
                center_y += bbox[axis][1]
            center_y = int(center_y / len(cluster[0]))
            for text, bbox in cluster[0]:
                bbox[axis][1] = center_y
                new_text_list.append(text)
                new_bbox_list.append(bbox)
            # print("cluster_list", cluster)
        return new_text_list, new_bbox_list

    def merge_textbox(textbox_list, in_objs):
        delete_obj = []
        threshold = 5
        for k in range(len(textbox_list)):
            tb1 = textbox_list[k]
            if tb1 not in in_objs and tb1 not in delete_obj:
                for m in range(k+1, len(textbox_list)):
                    tb2 = textbox_list[m]
                    if abs(tb1.bbox[1]-tb2.bbox[1]) <= threshold \
                            and abs(tb1.bbox[3]-tb2.bbox[3]) <= threshold:
                        if tb1.bbox[0] <= tb2.bbox[0]:
                            tb1.text = tb1.text + tb2.text
                        else:
                            tb1.text = tb2.text + tb1.text
                        tb1.bbox[0] = min(tb1.bbox[0], tb2.bbox[0])
                        tb1.bbox[2] = max(tb1.bbox[2], tb2.bbox[2])
                        delete_obj.append(tb2)
        for _obj in delete_obj:
            if _obj in textbox_list:
                textbox_list.remove(_obj)
        return textbox_list

    log("into image_preprocess")
    try:
        # 图片倾斜校正，写入原来的图片路径
        # print("image_process", image_path)
        g_r_i = get_rotated_image(image_np, image_path)
        if judge_error_code(g_r_i):
            if is_from_docx:
                return []
            else:
                return g_r_i

        image_np = cv2.imread(image_path)
        if image_np is None:
            return []

        # otr需要图片resize成模型所需大小, 写入另一个路径
        best_h, best_w = get_best_predict_size(image_np)
        image_resize = cv2.resize(image_np, (best_w, best_h), interpolation=cv2.INTER_AREA)
        image_resize_path = image_path.split(".")[0] + "_resize_otr." + image_path.split(".")[-1]
        cv2.imwrite(image_resize_path, image_resize)

        # 调用otr模型接口
        with open(image_resize_path, "rb") as f:
            image_bytes = f.read()
        list_line = from_otr_interface(image_bytes, is_from_pdf)
        if judge_error_code(list_line):
            return list_line

        # # 预处理
        # if is_from_pdf:
        #     prob = 0.2
        # else:
        #     prob = 0.5
        # with open(image_resize_path, "rb") as f:
        #     image_bytes = f.read()
        # img_new, inputs = table_preprocess(image_bytes, prob)
        # if type(img_new) is list and judge_error_code(img_new):
        #     return img_new
        # log("img_new.shape " + str(img_new.shape))
        #
        # # 调用模型运行接口
        # _dict = {"inputs": inputs, "md5": _global.get("md5")}
        # result = from_gpu_interface(_dict, model_type="otr", predictor_type="")
        # if judge_error_code(result):
        #     logging.error("from_gpu_interface failed! " + str(result))
        #     raise requests.exceptions.RequestException
        #
        # pred = result.get("preds")
        # gpu_time = result.get("gpu_time")
        # log("otr model predict time " + str(gpu_time))
        #
        # # # 解压numpy
        # # decompressed_array = io.BytesIO()
        # # decompressed_array.write(pred)
        # # decompressed_array.seek(0)
        # # pred = np.load(decompressed_array, allow_pickle=True)['arr_0']
        # # log("inputs.shape" + str(pred.shape))
        #
        # # 后处理
        # list_line = table_postprocess(img_new, pred, prob)
        # log("len(list_line) " + str(len(list_line)))
        # if judge_error_code(list_line):
        #     return list_line

        # otr resize后得到的bbox根据比例还原
        start_time = time.time()
        ratio = (image_np.shape[0]/best_h, image_np.shape[1]/best_w)
        for i in range(len(list_line)):
            point = list_line[i]
            list_line[i] = [int(point[0]*ratio[1]), int(point[1]*ratio[0]),
                            int(point[2]*ratio[1]), int(point[3]*ratio[0])]
        log("otr resize bbox recover " + str(time.time()-start_time))

        # ocr图片过大内存溢出，需resize
        start_time = time.time()
        threshold = 3000
        if image_np.shape[0] >= threshold or image_np.shape[1] >= threshold:
            best_h, best_w = get_best_predict_size2(image_np, threshold)
            image_resize = cv2.resize(image_np, (best_w, best_h), interpolation=cv2.INTER_AREA)
            # image_resize_path = image_path.split(".")[0] + "_resize_ocr." + image_path.split(".")[-1]
            # cv2.imwrite(image_resize_path, image_resize)
        log("ocr resize before " + str(time.time()-start_time))

        # 调用ocr模型接口
        with open(image_resize_path, "rb") as f:
            image_bytes = f.read()
        text_list, bbox_list = from_ocr_interface(image_bytes, is_table=True)
        if judge_error_code(text_list):
            return text_list

        # # PaddleOCR内部包括预处理，调用模型运行接口，后处理
        # paddle_ocr = PaddleOCR(use_angle_cls=True, lang="ch")
        # results = paddle_ocr.ocr(image_resize, det=True, rec=True, cls=True)
        # # 循环每张图片识别结果
        # text_list = []
        # bbox_list = []
        # for line in results:
        #     # print("ocr_interface line", line)
        #     text_list.append(line[-1][0])
        #     bbox_list.append(line[0])
        # if len(text_list) == 0:
        #     return []

        # ocr resize后的bbox还原
        ratio = (image_np.shape[0]/best_h, image_np.shape[1]/best_w)
        for i in range(len(bbox_list)):
            point = bbox_list[i]
            bbox_list[i] = [[int(point[0][0]*ratio[1]), int(point[0][1]*ratio[0])],
                            [int(point[1][0]*ratio[1]), int(point[1][1]*ratio[0])],
                            [int(point[2][0]*ratio[1]), int(point[2][1]*ratio[0])],
                            [int(point[3][0]*ratio[1]), int(point[3][1]*ratio[0])]]

        # 对文字框的y进行聚类
        text_list, bbox_list = get_cluster(text_list, bbox_list, 0)
        text_list, bbox_list = get_cluster(text_list, bbox_list, 2)

        # 调用现成方法形成表格
        try:
            from format_convert.convert_tree import TableLine
            list_lines = []
            for line in list_line:
                list_lines.append(LTLine(1, (line[0], line[1]), (line[2], line[3])))
            from format_convert.convert_tree import TextBox
            list_text_boxes = []
            for i in range(len(bbox_list)):
                bbox = bbox_list[i]
                b_text = text_list[i]
                list_text_boxes.append(TextBox([bbox[0][0], bbox[0][1],
                                                bbox[2][0], bbox[2][1]], b_text))

            lt = LineTable()
            tables, obj_in_table, _ = lt.recognize_table(list_text_boxes, list_lines, False)

            # 合并同一行textbox
            list_text_boxes = merge_textbox(list_text_boxes, obj_in_table)

            obj_list = []
            for table in tables:
                obj_list.append(_Table(table["table"], table["bbox"]))
            for text_box in list_text_boxes:
                if text_box not in obj_in_table:
                    obj_list.append(_Sentence(text_box.get_text(), text_box.bbox))
            return obj_list
        except:
            traceback.print_exc()
            return [-8]

    except Exception as e:
        log("image_preprocess error")
        traceback.print_exc()
        return [-1]


@memory_decorator
def picture2text(path, html=False):
    log("into picture2text")
    try:
        # 判断图片中表格
        img = cv2.imread(path)
        if img is None:
            return [-3]

        text = image_process(img, path)
        if judge_error_code(text):
            return text

        if html:
            text = add_div(text)
        return [text]
    except Exception as e:
        log("picture2text error!")
        print("picture2text", traceback.print_exc())
        return [-1]


def get_best_predict_size(image_np, times=64):
    sizes = []
    for i in range(1, 100):
        if i*times <= 1300:
            sizes.append(i*times)
    sizes.sort(key=lambda x: x, reverse=True)

    min_len = 10000
    best_height = sizes[0]
    for height in sizes:
        if abs(image_np.shape[0] - height) < min_len:
            min_len = abs(image_np.shape[0] - height)
            best_height = height

    min_len = 10000
    best_width = sizes[0]
    for width in sizes:
        if abs(image_np.shape[1] - width) < min_len:
            min_len = abs(image_np.shape[1] - width)
            best_width = width

    return best_height, best_width


def get_best_predict_size2(image_np, threshold=3000):
    h, w = image_np.shape[:2]
    scale = threshold / max(h, w)
    h = int(h * scale)
    w = int(w * scale)
    return h, w


class ImageConvert:
    def __init__(self, path, unique_type_dir):
        from format_convert.convert_tree import _Document
        self._doc = _Document(path)
        self.path = path
        self.unique_type_dir = unique_type_dir

    def init_package(self):
        # 各个包初始化
        try:
            with open(self.path, "rb") as f:
                self.image = f.read()
        except:
            log("cannot open image!")
            traceback.print_exc()
            self._doc.error_code = [-3]

    def convert(self):
        from format_convert.convert_tree import _Page, _Image
        self.init_package()
        if self._doc.error_code is not None:
            return

        _page = _Page(None, 0)
        _image = _Image(self.image, self.path)
        _page.add_child(_image)
        self._doc.add_child(_page)

    def get_html(self):
        try:
            self.convert()
        except:
            traceback.print_exc()
            self._doc.error_code = [-1]
        if self._doc.error_code is not None:
            return self._doc.error_code
        return self._doc.get_html()