import os import sys from PIL import Image __dir__ = os.path.dirname(os.path.abspath(__file__)) sys.path.append(__dir__) sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '../..'))) os.environ["FLAGS_allocator_strategy"] = 'auto_growth' import cv2 import numpy as np import math import time import traceback import paddle import tools.infer.utility as utility from ppocr.postprocess import build_post_process from ppocr.utils.logging import get_logger from ppocr.utils.utility import get_image_file_list, check_and_read_gif logger = get_logger() class TextRecognizer(object): def __init__(self, args): self.rec_image_shape = [int(v) for v in args.rec_image_shape.split(",")] self.rec_batch_num = args.rec_batch_num postprocess_params = { 'name': 'CTCLabelDecode', "character_dict_path": args.rec_char_dict_path, "use_space_char": args.use_space_char } self.postprocess_op = build_post_process(postprocess_params) self.predictor, self.input_tensor, self.output_tensors, self.config = \ utility.create_predictor(args, 'rec', logger) def resize_norm_img_section(self, img, max_wh_ratio): # print("rec resize for section") imgC, imgH, imgW = self.rec_image_shape assert imgC == img.shape[2] rec_precision_level = os.environ.get("OCR_REC_PRECISION") max_w = imgH * 48 # max_w = 2304 if rec_precision_level =='0': imgW = max_w elif rec_precision_level == '1': imgW = int((imgH * max_wh_ratio)) if imgW <= max_w / 2: imgW = max_w / 2 else: imgW = max_w elif rec_precision_level == '2': imgW = int((imgH * max_wh_ratio)) if imgW <= max_w / 4: imgW = max_w / 4 elif imgW > max_w / 4 and imgW <= max_w / 2: imgW = max_w / 2 elif imgW > max_w / 2 and imgW <= 3 * max_w / 4: imgW = 3 * max_w / 4 else: imgW = max_w else: imgW = int((imgH * max_wh_ratio)) if imgW <= max_w / 6: imgW = max_w / 6 elif imgW > max_w / 6 and imgW <= max_w / 3: imgW = max_w / 3 elif imgW > max_w / 3 and imgW <= max_w / 2: imgW = max_w / 2 elif imgW > max_w / 2 and imgW <= 2 * max_w / 3: imgW = 2 * max_w / 3 elif imgW > 2 *max_w / 3 and imgW <= 5 * max_w / 6: imgW = 5 * max_w / 6 else: imgW = max_w imgW = int(imgW) h, w = img.shape[:2] ratio = w / float(h) if math.ceil(imgH * ratio) > imgW: resized_w = imgW else: resized_w = int(math.ceil(imgH * ratio)) resized_image = cv2.resize(img, (resized_w, imgH)) resized_image = resized_image.astype('float32') resized_image = resized_image.transpose((2, 0, 1)) / 255 resized_image -= 0.5 resized_image /= 0.5 padding_im = np.zeros((imgC, imgH, imgW), dtype=np.float32) padding_im[:, :, 0:resized_w] = resized_image return padding_im def __call__(self, img_list): img_num = len(img_list) # Calculate the aspect ratio of all text bars width_list = [] for img in img_list: width_list.append(img.shape[1] / float(img.shape[0])) # Sorting can speed up the recognition process indices = np.argsort(np.array(width_list)) rec_res = [['', 0.0]] * img_num if img_num <= 0: return rec_res, 0 max_batnum = 24 min_batnum = 8 if os.environ.get("OCR_REC_MAX_BATNUM") is not None: max_batnum = int(os.environ.get("OCR_REC_MAX_BATNUM")) if os.environ.get("OCR_REC_MIN_BATNUM") is not None: min_batnum = int(os.environ.get("OCR_REC_MIN_BATNUM")) assert max_batnum / min_batnum == int(max_batnum / min_batnum), "max_batnum must be multiple of min_batnum." img_num_left = img_num img_no_count = 0 st = time.time() if img_num_left > max_batnum: batch_num = max_batnum batch_num = int(batch_num) for beg_img_no in range(img_no_count, int(img_num_left / batch_num) * batch_num, batch_num): end_img_no = beg_img_no + batch_num norm_img_batch = [] max_wh_ratio = 0 for ino in range(beg_img_no, end_img_no): h, w = img_list[indices[ino]].shape[0:2] wh_ratio = w * 1.0 / h max_wh_ratio = max(max_wh_ratio, wh_ratio) for ino in range(beg_img_no, end_img_no): norm_img = self.resize_norm_img_section(img_list[indices[ino]], max_wh_ratio) norm_img = norm_img[np.newaxis, :] norm_img_batch.append(norm_img) norm_img_batch = np.concatenate(norm_img_batch, axis=0) norm_img_batch = norm_img_batch.copy() self.input_tensor.copy_from_cpu(norm_img_batch) self.predictor.run() outputs = [] for output_tensor in self.output_tensors: output = output_tensor.copy_to_cpu() outputs.append(output) if len(outputs) != 1: preds = outputs else: preds = outputs[0] rec_result = self.postprocess_op(preds) for rno in range(len(rec_result)): rec_res[indices[beg_img_no + rno]] = rec_result[rno] img_no_count = int(img_num_left / batch_num) * batch_num img_num_left = img_num_left - int(img_num_left / batch_num) * batch_num batch_num = math.ceil(img_num_left / min_batnum) * min_batnum batch_num = int(batch_num) Dnum = batch_num - img_num_left for dno in range(Dnum): indices = np.append(indices,img_num + dno) rec_res.append(['', 0.0]) beg_img_no = img_no_count end_img_no = img_num norm_img_batch = [] max_wh_ratio = 0 for ino in range(beg_img_no, end_img_no): h, w = img_list[indices[ino]].shape[0:2] wh_ratio = w * 1.0 / h max_wh_ratio = max(max_wh_ratio, wh_ratio) for ino in range(beg_img_no, end_img_no): norm_img = self.resize_norm_img_section(img_list[indices[ino]], max_wh_ratio) norm_img = norm_img[np.newaxis, :] norm_img_batch.append(norm_img) norm_img_batch = np.concatenate(norm_img_batch) if norm_img_batch.shape[0] != batch_num: img_tmp = np.zeros((batch_num - norm_img_batch.shape[0], norm_img_batch.shape[1], norm_img_batch.shape[2], norm_img_batch.shape[3]), dtype=np.float32) norm_img_batch = np.concatenate([norm_img_batch, img_tmp]) norm_img_batch = norm_img_batch.copy() self.input_tensor.copy_from_cpu(norm_img_batch) self.predictor.run() outputs = [] for output_tensor in self.output_tensors: output = output_tensor.copy_to_cpu() outputs.append(output) if len(outputs) != 1: preds = outputs else: preds = outputs[0] rec_result = self.postprocess_op(preds) for rno in range(len(rec_result)): rec_res[indices[beg_img_no + rno]] = rec_result[rno] return rec_res, time.time() - st def main(args): image_file_list = get_image_file_list(args.image_dir) text_recognizer = TextRecognizer(args) valid_image_file_list = [] img_list = [] logger.info( "In PP-OCRv3, rec_image_shape parameter defaults to '3, 48, 320', " "if you are using recognition model with PP-OCRv2 or an older version, please set --rec_image_shape='3,32,320" ) # warmup 2 times if args.warmup: img = np.random.uniform(0, 255, [48, 320, 3]).astype(np.uint8) for i in range(2): res = text_recognizer([img] * int(args.rec_batch_num)) for image_file in image_file_list: img, flag = check_and_read_gif(image_file) if not flag: img = cv2.imread(image_file) if img is None: logger.info("error in loading image:{}".format(image_file)) continue valid_image_file_list.append(image_file) img_list.append(img) try: rec_res, _ = text_recognizer(img_list) except Exception as E: logger.info(traceback.format_exc()) logger.info(E) exit() for ino in range(len(img_list)): logger.info("Predicts of {}:{}".format(valid_image_file_list[ino], rec_res[ino])) if args.benchmark: text_recognizer.autolog.report() if __name__ == "__main__": main(utility.parse_args())