VerificationCode/chinese_recognize/inference_char.py

import os
import sys
import tensorflow as tf
import cv2
import numpy as np
sys.path.append(os.path.dirname(os.path.abspath(__file__)) + "/../")
from chinese_recognize.model import cnn_net
from utils import pil_resize

package_dir = os.path.abspath(os.path.dirname(__file__))
image_shape = (40, 40, 3)
model_path = package_dir + "/models/char_f1_0.93.h5"
with open(package_dir + "/chinese_5710.txt") as f:
    char_str = f.read()


def recognize(image_np_list, model=None, sess=None):
    if sess is None:
        sess = tf.compat.v1.Session(graph=tf.Graph())

    if model is None:
        with sess.as_default():
            with sess.graph.as_default():
                model = cnn_net(input_shape=image_shape)
                model.load_weights(model_path)

    if len(image_np_list) > 30:
        raise

    # 准备批次数据
    X = np.zeros((len(image_np_list), image_shape[0], image_shape[1], image_shape[2]))
    for i in range(len(image_np_list)):
        img = pil_resize(image_np_list[i], image_shape[0], image_shape[1])
        img = img / 255.
        X[i] = img

    with sess.as_default():
        with sess.graph.as_default():
            pred = model.predict(X)
    index_list = np.argmax(pred, axis=1).tolist()

    char_list = []
    for index in index_list:
        char_list.append(char_str[index])
    return char_list


if __name__ == "__main__":
    _path = "D:/Project/captcha/data/test/char_2.jpg"
    print(recognize([cv2.imread(_path)]))