search

intel / handwritten-chinese-ocr-samples

End-to-end model training and deployment reference for handwritten Chinese text recognition, and can also be extended to other languages.
Other
145 stars 31 forks source link

About the result of the example img 000000.jpg #7

Closed Randy-1009 closed 2 years ago

Randy-1009 commented 2 years ago

Why the last several words can not be recognized? It's because the program self or some length limit? 000000

Randy-1009 commented 2 years ago

The result is "仍有“铁马冰河入梦来”的情怀，一生心忧天", where is "下，被世人所"

bliu3650 commented 2 years ago

Why the last several words can not be recognized? It's because the program self or some length limit? 000000

You're right. Our previous released models (for SCUT-EPT dataset) had a limitation on the length of input, because of the deployment tool (openvino) which required the fixed length.

For this 000000.jpg sample, with 1318x48 resolution, we firstly resized it to 2636x96, and ignored the last 636 (=2636-2000) columns. So in the prediction, you can not see that "下，被世人所". But of course, you can remove that limitation by regenerating the model (bin/xml) and setting length as long as you need for deployment with OpenVINO.

Randy-1009 commented 2 years ago

Thank you!