def single_recognition(img, model_dir):
img_w = 256
img_h = 32
img_c = 1
downsample_factor = 4
if img_c == 1:
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = cv2.resize(img, (img_w, img_h))
img = np.expand_dims(img, axis=-1)
img = img / 255.0 * 2.0 - 1.0
img_batch = np.zeros((1, img_h, img_w, img_c))
img_batch[0, :, :, :] = img
# print(gray_img.shape)
model_for_predict = model(is_training=False, img_shape=(img_h, img_w, img_c), num_classes=11, max_label_length=26)
# model_for_predict = vgg_b_ctc.model(is_training=False, img_size=(256,32), num_classes=11, max_label_length=25)
model_for_predict.load_weights(model_dir)
y_pred_probMatrix = model_for_predict.predict(img_batch)
# Decode 阶段
y_pred_labels_tensor_list, _ = keras.backend.ctc_decode(y_pred_probMatrix, [img_w // downsample_factor],
greedy=True) # 使用的是最简单的贪婪算法
y_pred_labels_tensor = y_pred_labels_tensor_list[0]
y_pred_labels = keras.backend.get_value(y_pred_labels_tensor) # 现在还是字符编码
# 转换成字符串
y_pred_text = ''
for num in y_pred_labels[0]:
y_pred_text += num2char_dict[num]
# print(y_pred_labels)
return y_pred_text
def single_recognition(img, w=256, h=32, debug=False):
sess = K.get_session()
keras.backend.clear_session()
img_w = 256
downsample_factor = 4
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
gray_img = cv2.resize(img, (w, h))
gray_img = np.expand_dims(gray_img, axis=-1)
gray_img = gray_img / 255.0 * 2.0 - 1.0
img_batch = np.zeros((1, h, w, 1))
img_batch[0, :, :, :] = gray_img
model_for_predict = model(is_training=False,
img_size=(256, 32),
num_classes=11,
max_label_length=26)
model_for_predict.load_weights('crnn/model/ocr-bankcard-keras-new.h5')
y_pred_probMatrix = model_for_predict.predict(img_batch)
# Decode 阶段
y_pred_labels_tensor_list, _ = keras.backend.ctc_decode(
y_pred_probMatrix, [img_w // downsample_factor],
greedy=True) # 使用的是最简单的贪婪算法
y_pred_labels_tensor = y_pred_labels_tensor_list[0]
y_pred_labels = keras.backend.get_value(y_pred_labels_tensor) # 现在还是字符编码
# 转换成字符串
y_pred_text = ''
for num in y_pred_labels[0]:
y_pred_text += num2char_dict[num]
# print(y_pred_labels)
K.set_session(sess)
return y_pred_text
def main():
img_path_list = os.listdir(test_dir)
img_path_list = sorted(img_path_list)
for num, img_name in enumerate(img_path_list):
img = Image(cv2.imread(test_dir + img_name))
cv2.imwrite(res_dir + 'card_' + str(num + 1) + '.jpg', img.pos_img)
print(img_name + '定位完成')
pre_model = model(is_training=False,
img_size=(256, 32),
num_classes=11,
max_label_length=26)
res = PredictLabels_by_filename(pre_model,
res_dir, (256, 32),
downsample_factor=4,
weight_path='model/train_weight.h5')
res_key = sorted(res)
f = open(res_dir + 'result.txt', 'w')
for img_name in res_key:
f.write(img_name[:-4] + ':' + res[img_name] + '\n')
f.close()
def main():
'''
model_choice: 0--vgg_bgru_ctc, 1--vgg_blstm_ctc, 2--resnet_blstm_ctc 3--resnet18_blstm
'''
model_dict = {
0: "vgg_bgru_ctc",
1: "vgg_blstm_ctc",
2: "resnet_blstm_ctc",
3: 'resnet18_blstm'
}
model_choice = 1
model_for_train, model_for_predict = None, None
# 各种路径 以及参数
current_time = time.strftime('%Y_%m_%d_%H_%M_%S',
time.localtime(time.time()))
weight_save_path = "../trained_weights/{}_{}_best_weight.h5".format(
current_time, model_dict[model_choice])
# weight_save_path = "../trained_weights/2018_12_16_14_25_51_vgg_bgru_ctc_best_weight.h5"
# 数字训练路径
# img_data_dir = "../data/numbers_croped"
# train_txt_path = "../data/data_txt/numbers/numbers_train.txt"
# val_txt_path = "../data/data_txt/numbers/numbers_val.txt"
# img_size = (128, 32) # W*H
# 各种训练时候的参数
# num_classes = 11 # 包含“blank”
# max_label_length = 12
# downsample_factor = 4
# epochs = 100
# img_data_dir = "../data/all_data_croped"
# train_txt_path = "../data/data_txt/all_except_long/train_data.txt"
# val_txt_path = "../data/data_txt/all_except_long/test_data.txt"
# # 300w+ 训练路径
img_data_dir = "../data/img_300w/img"
train_txt_path = "../data/img_300w/txt/train.txt"
val_txt_path = "../data/img_300w/txt/test.txt"
# part 300w+ 训练参数
#路径
# img_data_dir = "../data/part_300w/img"
# train_txt_path = "../data/part_300w/txt/train.txt"
# val_txt_path = "../data/part_300w/txt/test.txt"
# 训练参数
img_size = (280, 32) # W*H
num_classes = 5991 # 把最后以为当成"blank"， 舍弃掉第一位
max_label_length = 10
downsample_factor = 4
epochs = 7
if model_choice == 0:
# vgg_bgru_ctc
model_for_train = vgg_bgru_ctc.model(is_training=True,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
model_for_predict = vgg_bgru_ctc.model(
is_training=False,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
elif model_choice == 1:
# vgg_blstm_ctc
model_for_train = vgg_blstm_ctc.model(
is_training=True,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
model_for_predict = vgg_blstm_ctc.model(
is_training=False,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
elif model_choice == 2:
# resnet_blstm_ctc
model_for_train = resnet_bgru_ctc.model(
is_training=True,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
model_for_predict = resnet_bgru_ctc.model(
is_training=False,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
elif model_choice == 3:
# resnet18_blstm
model_for_train = resnet18_blstm.model(
is_training=True,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
model_for_predict = resnet18_blstm.model(
is_training=False,
img_size=img_size,
num_classes=num_classes,
max_label_length=max_label_length)
# 训练模型
train_model(model_for_train,
img_data_dir,
train_txt_path,
val_txt_path,
weight_save_path,
epochs=epochs,
img_size=img_size,
batch_size=512,
max_label_length=max_label_length,
down_sample_factor=downsample_factor)
# 使用模型进行预测
predict_labels = PredictLabels(model_for_predict,
img_data_dir,
val_txt_path,
img_size,
downsample_factor,
batch_size=128,
weight_path=weight_save_path)
# check accuracy
acc, misclassified = check_acc(predict_labels, val_txt_path)
print("accuracy on the on the val_data is {}.".format(acc))
# 保存预测的结果
test_result_save_path = "../predicted_results/{}_acc_{:.2f}_{}_result.txt".format(
current_time, acc, model_dict[model_choice])
result_txt = open(test_result_save_path, 'w')
result_txt.write("acc :{}.\n".format(acc))
result_txt.write("misclassified: {}.\n".format(len(misclassified)))
for key, value in misclassified.items():
result_txt.write(str(key) + ": " + str(value) + '\n')
result_txt.write("all results:\n")
for key, value in predict_labels.items():
result_txt.write(str(key) + ": " + str(value) + '\n')
result_txt.close()
return 0