| # pip install numpy==1.26.4 opencv-python==4.6.0.66 | |
| # see doc\lang\programming\pytorch\文本检测\DBNET 论文代码都有 | |
| """ | |
| 给 DBNet 官方代码用 | |
| 将阿里OCR 的识别结果（图片和标注）转换成 icdar2015 格式 (注意：它的文本是含 utf8 bom 的) | |
| """ | |
| """ | |
| icdar2015 文本检测数据集 | |
| 标注格式: x1,y1,x2,y2,x3,y3,x4,y4,text | |
| 其中, x1,y1为左上角坐标,x2,y2为右上角坐标,x3,y3为右下角坐标,x4,y4为左下角坐标。 | |
| ### 表示text难以辨认。 | |
| """ | |
| import random | |
| from pathlib import Path | |
| import os | |
| import glob | |
| import base64 | |
| from importlib.resources import path | |
| import math | |
| import numpy as np | |
| import cv2 | |
| import json | |
| import decimal | |
| import datetime | |
| from pickletools import uint8 | |
| class DecimalEncoder(json.JSONEncoder): | |
| def default(self, o): | |
| if isinstance(o, decimal.Decimal): | |
| return float(o) | |
| elif isinstance(o, datetime.datetime): | |
| return str(o) | |
| super(DecimalEncoder, self).default(o) | |
| def save_json(filename, dics): | |
| with open(filename, 'w', encoding='utf-8') as fp: | |
| json.dump(dics, fp, indent=4, cls=DecimalEncoder, ensure_ascii=False) | |
| fp.close() | |
| def load_json(filename): | |
| with open(filename, encoding='utf-8') as fp: | |
| js = json.load(fp) | |
| fp.close() | |
| return js | |
| # convert string to json | |
| def parse(s): | |
| return json.loads(s, strict=False) | |
| # convert dict to string | |
| def string(d): | |
| return json.dumps(d, cls=DecimalEncoder, ensure_ascii=False) | |
| def transform(points, M): | |
| # points 算出四个点变换后移动到哪里了 | |
| # points = np.array([[word_x, word_y], # 左上 | |
| # [word_x + word_width, word_y], # 右上 | |
| # [word_x + word_width, word_y + word_height], # 右下 | |
| # [word_x, word_y + word_height], # 左下 | |
| # ]) | |
| # add ones | |
| ones = np.ones(shape=(len(points), 1)) | |
| points_ones = np.hstack([points, ones]) | |
| # transform points | |
| transformed_points = M.dot(points_ones.T).T | |
| transformed_points_int = np.round( | |
| transformed_points, decimals=0).astype(np.int32) # 批量四舍五入 | |
| return transformed_points_int | |
| def cutPoly(img, pts): | |
| # img = cv2.imdecode(np.fromfile('./t.png', dtype=np.uint8), -1) | |
| # pts = np.array([[10,150],[150,100],[300,150],[350,100],[310,20],[35,10]]) | |
| ## (1) Crop the bounding rect | |
| rect = cv2.boundingRect(pts) | |
| x,y,w,h = rect | |
| croped = img[y:y+h, x:x+w].copy() | |
| ## (2) make mask | |
| pts = pts - pts.min(axis=0) | |
| mask = np.zeros(croped.shape[:2], np.uint8) | |
| cv2.drawContours(mask, [pts], -1, (255, 255, 255), -1, cv2.LINE_AA) | |
| ## (3) do bit-op | |
| dst = cv2.bitwise_and(croped, croped, mask=mask) | |
| ## (4) add the white background | |
| bg = np.ones_like(croped, np.uint8)*255 | |
| cv2.bitwise_not(bg,bg, mask=mask) | |
| dst2 = bg+ dst | |
| # cv2.imwrite("croped.png", croped) | |
| # cv2.imwrite("mask.png", mask) | |
| # cv2.imwrite("dst.png", dst) | |
| # cv2.imwrite("dst2.png", dst2) | |
| return dst2 | |
| if __name__ == "__main__": | |
| # 验证原版的文本标记框 | |
| # im = './datasets/icdar2015/train_images/img_1.jpg' | |
| # gt = './datasets/icdar2015/train_gts/gt_img_1.txt' | |
| # 验证自已生成的标记框 | |
| im = './icdar2015_aliocr/train_images/img_000001.jpg' | |
| gt = './icdar2015_aliocr/train_gts/gt_img_000001.txt' | |
| if os.path.exists(gt): | |
| items = [] | |
| reader = open(gt, 'r', encoding='utf-8-sig').readlines() | |
| for line in reader: | |
| item = {} | |
| parts = line.strip().split(',') | |
| label = parts[-1] | |
| if 'TD' in gt and label == '1': | |
| label = '###' | |
| line = [i.strip('\ufeff').strip('\xef\xbb\xbf') for i in parts] | |
| if 'icdar' in gt: | |
| poly = np.array(list(map(float, line[:8]))).reshape( | |
| (-1, 2)).tolist() | |
| else: | |
| num_points = math.floor((len(line) - 1) / 2) * 2 | |
| poly = np.array(list(map(float, line[:num_points]))).reshape( | |
| (-1, 2)).tolist() | |
| item['poly'] = poly | |
| item['text'] = label | |
| # 多边形是用一个个的点表示的，起点连接第二个点，第二个连接第三个 ... 最后一点连接起点，构成一个闭合的区域 | |
| item['points'] = poly | |
| # 此标记表示文字模糊不可辨认，文本框的标记是不可靠的 | |
| item['ignore'] = True if label == '###' else False | |
| items.append(item) | |
| img = cv2.imdecode(np.fromfile(im, dtype=np.uint8), -1) | |
| # DBNet 原版代码只能处理彩图，所以统一处理成彩图 | |
| img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) | |
| for i in range(len(items)): | |
| poly = items[i]['poly'] | |
| poly = np.array(poly) | |
| poly = poly.astype(np.int32) | |
| # cv2.fillPoly(img, pts=[ poly ], color=(0, 0, 255)) | |
| b = random.randint(0, 255) # 用来生成[a,b]之间的随意整数，包括两个边界值。 | |
| g = random.randint(0, 255) | |
| r = random.randint(0, 255) | |
| # 只画线，不填充 # 就是画线，从起点连到第二个点 ... 最后一个点连到第一个点 | |
| cv2.polylines(img, [poly], isClosed=True, | |
| color=(b, g, r), thickness=1) | |
| cv2.imwrite("poly.jpg", img) | |
| # cv2.imshow("poly", img) | |
| # cv2.waitKey() | |
| # 开始转换 | |
| out_dir = 'icdar2015_aliocr' | |
| if os.path.exists(out_dir): | |
| import shutil | |
| shutil.rmtree(out_dir) | |
| # https://help.aliyun.com/document_detail/294540.html 阿里云ocr结果字段定义 | |
| # prism-wordsInfo 里的 angle 文字块的角度，这个角度只影响width和height，当角度为-90、90、-270、270，width和height的值需要自行互换 | |
| dir_json = './data/json' # '/yingedu/www/ocr_server/data/json' | |
| dir_img = './data/img' # '/yingedu/www/ocr_server/data/img' | |
| train_list = [] | |
| train_list_path = os.path.join(out_dir, 'train_list.txt') | |
| test_list = [] | |
| test_list_path = os.path.join(out_dir, 'test_list.txt') | |
| g_count = 1 | |
| json_paths = glob.glob('{}/*.json'.format(dir_json), recursive=True) | |
| for json_path in json_paths: | |
| base = Path(json_path).stem | |
| img_train_path = os.path.join(dir_img, '{}.txt'.format(base)) | |
| if not os.path.exists(img_train_path): # 没有相应的图片，可能被删除了 | |
| continue | |
| jsn = load_json(json_path) | |
| with open(img_train_path, "r", encoding="utf-8") as fp: | |
| imgdata = fp.read() | |
| imgdata = base64.b64decode(imgdata) | |
| imgdata = np.frombuffer(imgdata, np.uint8) | |
| img = cv2.imdecode(imgdata, cv2.IMREAD_UNCHANGED) | |
| # cv2.imshow('img', img) | |
| # cv2.waitKey(0) | |
| if len(img.shape) != 3: # 转彩图 | |
| img_color = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) | |
| img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) # DBNet 原版只能处理彩图，这里转一下 | |
| else: | |
| img_color = img.copy() | |
| img_color_origin = img_color.copy() | |
| img_color_origin2 = img_color.copy() | |
| img_name = "img_{:06d}.jpg".format(g_count) | |
| gt_name = "gt_img_{:06d}.txt".format(g_count) | |
| is_train_img = random.choices([0, 1], weights=[0.15, 0.85])[0] | |
| # 85% 的概率是训练图 | |
| gt_txt_list = [] | |
| img_train_path = os.path.join(out_dir, 'train_images', img_name) | |
| img_train_gt_path = os.path.join(out_dir, 'train_gts', gt_name) | |
| img_test_path = os.path.join(out_dir, 'test_images', img_name) | |
| img_test_gt_path = os.path.join(out_dir, 'test_gts', gt_name) | |
| dir1 = os.path.dirname(img_train_path) | |
| dir2 = os.path.dirname(img_train_gt_path) | |
| dir3 = os.path.dirname(img_test_path) | |
| dir4 = os.path.dirname(img_test_gt_path) | |
| if not os.path.exists(dir1): | |
| os.makedirs(dir1) | |
| if not os.path.exists(dir2): | |
| os.makedirs(dir2) | |
| if not os.path.exists(dir3): | |
| os.makedirs(dir3) | |
| if not os.path.exists(dir4): | |
| os.makedirs(dir4) | |
| if is_train_img: | |
| train_list.append(img_name) | |
| cv2.imwrite(img_train_path, img) | |
| else: | |
| test_list.append(img_name) | |
| cv2.imwrite(img_test_path, img) | |
| wordsInfo = jsn['prism_wordsInfo'] | |
| for j in range(len(wordsInfo)): | |
| jo = wordsInfo[j] | |
| word = jo["word"] | |
| # prism-wordsInfo 里的 angle 文字块的角度，这个角度只影响width和height，当角度为-90、90、-270、270，width和height的值需要自行互换 | |
| angle = jo['angle'] | |
| img_color = img_color_origin.copy() | |
| """ | |
| x y 宽高全部不靠谱, pos 里是对的 | |
| """ | |
| # word_x = jo['x'] | |
| # word_y = jo['y'] | |
| # word_width = jo['width'] | |
| # word_height = jo['height'] | |
| # if abs(angle) == 90 or abs(angle) == 270: | |
| # word_width = jo['height'] | |
| # word_height = jo['width'] | |
| # elif angle != 0: | |
| # # 变换前画出绿框，方便追踪点的前后变化 | |
| # img_color = cv2.rectangle(img_color, (word_x, word_y), (word_x + word_width, word_y + word_height), (0, 255, 0), 2) # 矩形的左上角, 矩形的右下角 | |
| # cv2.imshow("green", img_color) | |
| # cv2.waitKey(0) | |
| # # 变换前的多边形蓝框 | |
| # points = np.array([ | |
| # [word_x, word_y], # 左上 | |
| # [word_x + word_width, word_y], # 右上 | |
| # [word_x + word_width, word_y + word_height], # 右下 | |
| # [word_x, word_y + word_height], # 左下 | |
| # ]) | |
| # # cv2.fillPoly(img_color, pts=[points], color=(255, 0, 0)) # 填充 | |
| # cv2.polylines(img_color, [points], isClosed=True, color=( | |
| # 255, 0, 0), thickness=1) # 只画线，不填充 | |
| # cv2.imshow("polys", img_color) | |
| # cv2.waitKey(0) | |
| # # 获取图像的维度，并计算中心 | |
| # (h, w) = img_color.shape[:2] | |
| # (cX, cY) = (w // 2, h // 2) | |
| # # - (cX,cY): 旋转的中心点坐标 | |
| # # - 180: 旋转的度数，正度数表示逆时针旋转，而负度数表示顺时针旋转。 | |
| # # - 1.0：旋转后图像的大小，1.0原图，2.0变成原来的2倍，0.5变成原来的0.5倍 | |
| # # 1° = π/180弧度 1 弧度 = 180 / 3.1415926 // 0.0190033 是Mathematica 算出来的弧度，先转换成角度 // -0.0190033 * (180 / 3.1415926) | |
| # M = cv2.getRotationMatrix2D((cX, cY), angle, 1.0) | |
| # img_color = cv2.warpAffine(img_color, M, (w, h)) | |
| # img_color_transform = img_color.copy() | |
| # cv2.imshow("after trans", img_color) | |
| # cv2.waitKey(0) | |
| # # https://docs.opencv.org/2.4/doc/tutorials/imgproc/imgtrans/warp_affine/warp_affine.html # 原理 | |
| # # https://stackoverflow.com/questions/30327659/how-can-i-remap-a-point-after-an-image-rotation # How can I remap a point after an image rotation? | |
| # # 如何得到移动后的坐标点 | |
| # # points 算出四个点变换后移动到哪里了 | |
| # points = np.array([[word_x, word_y], # 左上 | |
| # # 右上 | |
| # [word_x + word_width, word_y], | |
| # [word_x + word_width, word_y + \ | |
| # word_height], # 右下 | |
| # [word_x, word_y + word_height], # 左下 | |
| # ]) | |
| # # add ones | |
| # ones = np.ones(shape=(len(points), 1)) | |
| # points_ones = np.hstack([points, ones]) | |
| # # transform points | |
| # transformed_points = M.dot(points_ones.T).T | |
| # transformed_points_int = np.round( | |
| # transformed_points, decimals=0).astype(np.int32) # 批量四舍五入 | |
| # cv2.polylines(img_color, [transformed_points_int], isClosed=True, color=( | |
| # 0, 0, 255), thickness=2) # 画转换后的点 | |
| # cv2.polylines(img_color_origin, [points], isClosed=True, color=( | |
| # random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)), thickness=2) # 画转换前的点 | |
| # cv2.imshow("orgin", img_color_origin) | |
| # cv2.waitKey(0) | |
| # 四个角的位置 # 左上、右上、右下、左下，当NeedRotate为true时，如果最外层的angle不为0，需要按照angle矫正图片后，坐标才准确 | |
| pos = jo["pos"] | |
| x = int(pos[0]["x"]) # 左上 | |
| y = int(pos[0]["y"]) | |
| x2 = int(pos[2]["x"]) # 右下 | |
| y2 = int(pos[2]["y"]) | |
| lu = [pos[0]['x'], pos[0]['y']] # left up 四个角顺时针方向数 | |
| ru = [pos[1]['x'], pos[1]['y']] | |
| rd = [pos[2]['x'], pos[2]['y']] | |
| ld = [pos[3]['x'], pos[3]['y']] | |
| # 生成 icdar2015 格式的人工标记训练数据（用于训练官方DB） | |
| gt_txt_list.append( "{},{},{},{},{},{},{},{},{}".format(lu[0], lu[1], ru[0], ru[1], rd[0], rd[1], ld[0], ld[1], word) ) | |
| # 绘制矩形 | |
| start_point = (x, y) # 矩形的左上角 | |
| end_point = (x2, y2) # 矩形的右下角 | |
| color = (0, 0, 255) # BGR | |
| thickness = 2 | |
| # 逐行画框 | |
| img_color = cv2.rectangle(img_color, start_point, end_point, color, thickness) | |
| # cv2.imshow("box", img_color) | |
| # cv2.waitKey(0) | |
| gt_txt = "\n".join(gt_txt_list) | |
| if is_train_img: | |
| with open(img_train_gt_path, 'w', encoding='utf-8') as f: | |
| f.write(gt_txt) | |
| else: | |
| with open(img_test_gt_path, 'w', encoding='utf-8') as f: | |
| f.write(gt_txt) | |
| print(f'### one task one. {g_count} / {len(json_paths)}') | |
| g_count += 1 | |
| # points = [ lu, ru, rd, ld ] | |
| # points0 = np.array([[word_x, word_y], # 左上 | |
| # # 右上 | |
| # [word_x + word_width, word_y], | |
| # [word_x + word_width, word_y + \ | |
| # word_height], # 右下 | |
| # [word_x, word_y + word_height], # 左下 | |
| # ]) | |
| # points1 = np.array( [ lu, ru, rd, ld ] ) | |
| # if not (abs(angle) == 90 or abs(angle) == 270) and angle != 0: | |
| # points = transform( points, M ) | |
| # else: | |
| # points = np.array(points) | |
| # ps3 = np.array( | |
| # [ | |
| # [min( points[0][0], points1[0][0] ), min( points[0][1], points1[0][1] )], # 左上(取最两者中最小的) | |
| # [max( points[1][0], points1[1][0] ), min( points[1][1], points1[1][1] )], # 右上 | |
| # [max( points[2][0], points1[2][0] ), max( points[2][1], points1[2][1] )], # 右下 | |
| # [min( points[3][0], points1[3][0] ), max( points[3][1], points1[3][1] )] # 左下 | |
| # ] | |
| # ) | |
| # img_cuted = cutPoly(img, ps3) | |
| # cv2.imwrite(f'./tmp/{g_count}.jpg', img_cuted) | |
| # with open(f'./tmp/{g_count}.txt', 'w', encoding='utf-8') as f: | |
| # f.write(word) | |
| # g_count += 1 | |
| # cv2.polylines(img_color, [points], isClosed=True, color=( # 多边形，框得比较全 | |
| # 100, 0, 255), thickness=2) # 只画线，不填充 | |
| # cv2.polylines(img_color_origin, [ points1 ], isClosed=True, color=( | |
| # random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)), thickness=2) # 画转换前的点 | |
| # cv2.imshow("orgin", img_color_origin) | |
| # cv2.waitKey(0) | |
| # # cv2.imshow("box", img_color) | |
| # # cv2.waitKey(0) | |
| # # img_color = cv2.rectangle(img_color, points[0], points[2], color, thickness) # 正常矩形，框不完全 | |
| # # cv2.imshow("box", img_color) | |
| # # cv2.waitKey(0) | |
| # if not (abs(angle) == 90 or abs(angle) == 270) and angle != 0: | |
| # t = word | |
| # ps = np.array( | |
| # [ | |
| # [min( transformed_points_int[0][0], points[0][0] ), min( transformed_points_int[0][1], points[0][1] )], # 左上(取最两者中最小的) | |
| # [max( transformed_points_int[1][0], points[1][0] ), min( transformed_points_int[1][1], points[1][1] )], # 右上 | |
| # [max( transformed_points_int[2][0], points[2][0] ), max( transformed_points_int[2][1], points[2][1] )], # 右下 | |
| # [min( transformed_points_int[3][0], points[3][0] ), max( transformed_points_int[3][1], points[3][1] )] # 左下 | |
| # ] | |
| # ) | |
| # ps2 = np.array( | |
| # [ | |
| # [min( points0[0][0], points1[0][0] ), min( points0[0][1], points1[0][1] )], # 左上(取最两者中最小的) | |
| # [max( points0[1][0], points1[1][0] ), min( points0[1][1], points1[1][1] )], # 右上 | |
| # [max( points0[2][0], points1[2][0] ), max( points0[2][1], points1[2][1] )], # 右下 | |
| # [min( points0[3][0], points1[3][0] ), max( points0[3][1], points1[3][1] )] # 左下 | |
| # ] | |
| # ) | |
| # # img_cuted = cutPoly(img_color_transform, ps) | |
| # # cv2.imwrite(f'./tmp/{g_count}.jpg', img_cuted) | |
| # # with open(f'./tmp/{g_count}.txt', 'w', encoding='utf-8') as f: | |
| # # f.write(word) | |
| # # g_count += 1 | |
| # cv2.polylines(img_color, [ ps ], isClosed=True, color=( | |
| # 255, 0, 0), thickness=2) # 只画线，不填充 | |
| # cv2.polylines(img_color_origin, [ ps2 ], isClosed=True, color=( | |
| # random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)), thickness=2) # 只画线，不填充 | |
| # cv2.imshow("orgin", img_color_origin) | |
| # cv2.waitKey(0) | |
| # img_cuted = cutPoly(img, ps2) | |
| # cv2.imwrite(f'./tmp/{g_count}.jpg', img_cuted) | |
| # with open(f'./tmp/{g_count}.txt', 'w', encoding='utf-8') as f: | |
| # f.write(word) | |
| # g_count += 1 | |
| # # cv2.imshow("box", img_color) | |
| # # cv2.waitKey(0) | |
| # lastx_mini = 0 # 下一个字符x 坐标的下界（肯定不小于这个值） | |
| # prew = 0 # 上一个字符的宽度 | |
| # words = "" | |
| # charInfo = jo["charInfo"] | |
| # min_cx = 9999 # 最小左上角 | |
| # min_cy = 9999 | |
| # max_cxcw = -1 # 最大右下角 | |
| # max_cych = -1 | |
| # for i in range(len(charInfo)): | |
| # joc = charInfo[i] | |
| # c = joc["word"] | |
| # cx = int(joc["x"]) | |
| # cy = int(joc["y"]) | |
| # cw = int(joc["w"]) | |
| # ch = int(joc["h"]) | |
| # if cx < min_cx: | |
| # min_cx = cx | |
| # if cy < min_cy: | |
| # min_cy = cy | |
| # if cx + cw > max_cxcw: | |
| # max_cxcw = cx + cw | |
| # if cy + ch > max_cych: | |
| # max_cych = cy + ch | |
| # # 绘制矩形 | |
| # start_point = (cx, cy) # 矩形的左上角 | |
| # end_point = (cx + cw, cy + ch) # 矩形的右下角 | |
| # color = (0, 0, 255) # BGR | |
| # thickness = 2 | |
| # # 逐字画框 | |
| # # img_color = cv2.rectangle( | |
| # # img_color, start_point, end_point, color, thickness) | |
| # # cv2.imshow("box", img_color) | |
| # # cv2.waitKey(0) | |
| # # 这个框更准一些 | |
| # # img_color = cv2.rectangle( | |
| # # img_color, (min_cx, min_cy), (max_cxcw, max_cych), (0, 255, 0), thickness) | |
| # # cv2.imshow("box", img_color) | |
| # # cv2.waitKey(0) | |
| # # fix me: 如果上面的行框的左边要比这里更左，那就以行框的左边为准 | |
| # # 因为发现单个字的框会有漏字的现想 | |
| # gt_txt_list.append("{},{},{},{},{},{},{},{},{}".format( | |
| # min_cx, min_cy, max_cxcw, min_cy, max_cxcw, max_cych, min_cx, max_cych, word)) | |
| # gt_txt = '\n'.join(gt_txt_list) | |
| # with open(img_gt_path, "w", encoding='utf-8-sig') as fp: | |
| # fp.write(gt_txt) | |
| train_list_txt = "\n".join(train_list) | |
| test_list_txt = "\n".join(test_list) | |
| with open(os.path.join(out_dir, "train_list.txt"), 'w', encoding='utf-8') as f: | |
| f.write(train_list_txt) | |
| with open(os.path.join(out_dir, "test_list.txt"), 'w', encoding='utf-8') as f: | |
| f.write(test_list_txt) | |
| print('### all task done.') | |