lyl8213/plate_recognition-lprnet

Tensorflow based. Use CNN and CTC loss. It is a light network for Plate Recognition

plate_recognition-lprnet's Issues

How to test single lpr plate .

Hi can you please write steps for training single lpr plate . your code seems to be for batch size 50.
I mean if i create a folder named test and put only one image . it should be able to give output.

how to train

请问 怎么训练自己的样本，我想训练自己的样本，但是不知道训练样本标签格式

Question: Output layer for graph freeze?

Hi, before all amazing work with Plate Recognition - LPRnet implementation.

I have been trying to freeze the graph but I can't figure out the output layer to use on freeze_graph.

This is th freezing code I'm trying to use.

python3 ../tensorflow/tensorflow/python/tools/freeze_graph.py \
		--input_graph=input.pb \
		--input_binary=false \
		--input_checkpoint=lprtf3.ckpt \
		--output_graph=output.pb \
		--output_node_name=??

The input.pb includes the result of the saves of

tf.train.write_graph(session.graph.as_graph_def(), MODEL_FOLDER, "input.pb")

Any ideas on what to use on output_node_name ??

Character segmenation

Can this network output the bounding box of each character in the image?

怎么测试

Full dataset

In the paper, it is mentioned that "In total, the dataset contains 11696 cropped license plate images, which are split as 9:1 into training and validation subsets respectively".
However, the number of images in the train folder is only 752.
Can you please share the full dataset?

Prediction script

Do you have a single prediction script that goes by single image
This kind of function will be extremely helpful to your project

Is possible to train two line number plate?

Hi @lyl8213 , thanks for your work.
I wonder can i train and recognize two line number plate with LPRnet ? e.g motorbike or number plate in other countries.

Error after train

If I put python LPRtf3.py and testthis is the output:

len(original_list) 50 len(detected_list) 1  test and detect length desn't match
time:0.7172296047210693
len(original_list) 50 len(detected_list) 1  test and detect length desn't match
time:0.7143244743347168
len(original_list) 50 len(detected_list) 1  test and detect length desn't match

What is it wrong?

transfer learning

is it possible to use checkpoints in your code for tranfser learning?

麻烦请问转pb模型的输出节点名称是什么呢？

如果想要转换成pb模型，那么该模型的输出节点名称是什么呢？在哪一层进行转换呢，我尝试都失败了，感谢！

The width and Height is misplaced

img_size = [94, 24]
inputs = tf.placeholder(tf.float32, shape=(b, img_size[0], img_size[1], num_channels))
however, the shape order should be [batchsize, height, width, channel]

识别时,相同词很大概率会当成合并为一个如何解决

如鲁AA1233变成 鲁A1233 ,目前输出是18长度,需要改大还是调空白符的阈值?有大神指导一下吗

有没有论文里面的数据集，我想训练一下

len(original_list) 50 len(detected_list) 1 test and detect length desn't match

shape inconsistency issue between input data and kernal

I saw in your code when image processing,you made input shape to NWHC format:
'''python
images = np.transpose(images, axes=[0, 2, 1, 3])
'''
but in network building ,the default kernal shape is NHWC format.

so I'm not quite sure if it's my understanding issue or it's a bug in your code.

1

the recognition speed

一张车牌识别需要60毫秒左右，怎么会和文章差距这么大

advantage of division and reduce_mean and square part layers

Hi @lyl8213
why we use this part of the code? what's the advantage of this part?

    cx = mean(sqrt(x))
    x = Lambda(lambda x: x[0]/x[1])([x,cx]) # division

    x1 = AveragePooling2D(pool_size=[4,1],strides=[4,1],padding='same')(inputs)
    cx1 = mean(sqrt(x1))
    x1 = Lambda(lambda x: x[0]/x[1])([x1,cx1]) # division

    x2 = AveragePooling2D(pool_size=[4,1],strides=[4,1],padding='same')(x2)
    cx2 = mean(sqrt(x2))
    x2 = Lambda(lambda x: x[0]/x[1])([x2,cx2]) # division

    x3 = AveragePooling2D(pool_size=[2,1],strides=[2,1],padding='same')(x3)
    cx3 = mean(sqrt(x3))
    x3 = Lambda(lambda x: x[0]/x[1])([x3,cx3]) # division

    x = concatenate([x,x1,x2,x3], axis=3)

可提供50w+中文车牌数据

车牌识别数据集,，包含多种车牌类型，各省份数量均匀且充足：qq1668486259

可提供数据支持

我可提供蓝色绿色黄色车牌数据以及相关代码，准确率可达95%+。q:1041357701

为什么训练到30epoch左右就出现train_cost为0，而且val_cost没有减少？而且使用model8.24best里的模型进行测试时都是出现长度不一致，得不到测试结果？

训练的现象如下：
Epoch 26/300, steps = 3822, train_cost = 0.000, logits_size = (24, 50, 67), val_cost = 12.472, val_ler = 0.146, time = 1.824s, learning_rate = 0.0009000000427477062 Epoch....... 26 Epoch 27/300, steps = 3969, train_cost = 0.000, logits_size = (24, 50, 67), val_cost = 13.910, val_ler = 0.148, time = 1.781s, learning_rate = 0.0009000000427477062 Epoch....... 27 Epoch 28/300, steps = 4116, train_cost = 0.000, logits_size = (24, 50, 67), val_cost = 13.581, val_ler = 0.156, time = 1.778s, learning_rate = 0.0008099999977275729 Epoch....... 28 Epoch 29/300, steps = 4263, train_cost = 0.000, logits_size = (24, 50, 67), val_cost = 13.747, val_ler = 0.156, time = 1.954s, learning_rate = 0.0008099999977275729

测试的现象如下：
len(original_list) 50 len(detected_list) 1 test and detect length desn't match time:0.2623302936553955 len(original_list) 50 len(detected_list) 1 test and detect length desn't match time:0.27623462677001953 len(original_list) 50 len(detected_list) 1 test and detect length desn't match time:0.2702765464782715 len(original_list) 50 len(detected_list) 1 test and detect length desn't match time:0.27227091789245605 len(original_list) 50 len(detected_list) 1 test and detect length desn't match

请问CTC对数概率log_prob可以转换为车牌的识别率么？

源文件LPRtf3.py的343行，log_prob有什么用？请问如何输出车牌的识别准确率呢？谢谢！
decoded, log_prob = tf.nn.ctc_beam_search_decoder(logits, seq_len, merge_repeated=False)

你训练的model acc有多少？

为什么达不到论文的95% ？

what are the training data label's last numbers?

hello, and thank you for your awesome code can you please explain what the last number in your training data label is? I mean the ones and zeros at the end.

UnicodeEncodeError

Epoch....... 33
Epoch 34/300, steps = 4998, train_cost = 0.000, train_ler = 0.000, val_cost = 1.423, val_ler = 0.023, time = 2.851s, learning_rate = 0.0008099999977275729
Epoch....... 34
time:0.9731936454772949
T/F: original(length) <-------> detectcted(length)
True Traceback (most recent call last):
File "LPRtf3.py", line 454, in
train(a)
File "LPRtf3.py", line 419, in train
c, steps = do_batch(train_gen,val_gen)
File "LPRtf3.py", line 406, in do_batch
do_report(val_gen,test_num)
File "LPRtf3.py", line 374, in do_report
report_accuracy(dd, test_targets)
File "LPRtf3.py", line 359, in report_accuracy
print(hit, number, "(", len(number), ") <-------> ", detect_number, "(", len(detect_number), ")")
UnicodeEncodeError: 'ascii' codec can't encode character '\u6d59' in position 2: ordinal not in range(128)

can the label_len be any length ？

I see before train model i should set the label_len = 7 #the length of plate character, but i dot see where use this label_len

车牌检测和车牌识别商用级别数据集

可提供一手车牌检测和车牌识别商用级别数据集，数据简介见链接：https://www.bilibili.com/video/bv1kU4y1e7Sb

how to train/test

train or test:test

UnboundLocalError Traceback (most recent call last)
in
444 if name == "main":
445 a = input('train or test:')
--> 446 train(a)

in train(a)
305 img_size=img_size,
306 num_channels=num_channels,
--> 307 label_len=label_len)
308
309 val_gen = TextImageGenerator(img_dir=vi,

in init(self, img_dir, label_file, batch_size, img_size, num_channels, label_len)
65 self.labels = []
66
---> 67 self.init()
68
69 def init(self):

i got this error every time i try to train or test
in init(self)
77 #else:
78 # label = dict[filename[:3]] + filename[4:10]
---> 79 label = encode_label(label)
80 self.labels.append(label)
81 self._num_examples += 1

UnboundLocalError: local variable 'label' referenced before assignment

if len(original_list) != len(detected_list):

it's like a bug, please fix it

Train on different type of licence plate?

Hi
Thank you for sharing your code, is it possible to use your code to train a different type of licence plates?

Here is a picture of the kind of number plate i want to train your network on :