search

potterhsu / SVHNClassifier

A TensorFlow implementation of Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks (http://arxiv.org/pdf/1312.6082.pdf)
GNU General Public License v3.0
205 stars 75 forks source link

Will it work for characters? #18

Open MarStarck opened 5 years ago

MarStarck commented 5 years ago

I wanna recognize images including short characters as well as digits, like "A25", "B03" and so on. I wonder could this structure work on this situation?

fkoorc commented 5 years ago

Looks like the training data only have digital, i test and characters is not working If you want character you have to add character training data

MarStarck commented 5 years ago

I have added character training data but the acc drops to about 70%

fkoorc commented 5 years ago

Actually my model only have digital and still the test accuracy is about 75% but if the picture is No distortion and very clear (maybe use some pre-process OCR) the accuracy can be about 93% I guess the model was overfitting. Or maybe your training data format had some problem may I ask your character data source?

MarStarck commented 5 years ago

the data source is synthetic text data by TextRecognitionDataGenerator I believe they are quite clear without distortion. So I think the 10-layer conv model cannot describe these characters, it's under fitting.

fkoorc commented 5 years ago

I guess you are right. I saw the data contains en/cn/fr/de? Looks quite complicate. If your data is large enough. I think 70% is quite good result If you still wanna promote accuracy, I suggest you can see these two paper

Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition（2014） Reading Text in the Wild with Convolutional Neural Networks（2014）

Where are you from? Maybe we can change email and discuss

MarStarck commented 5 years ago

@fkoorc I'm from China, my email: difei_coder@163.com