OCR

OCR System

• glyph recognition

• feature extraction

• segmentation

• language model

Paradigm Shift

LSTM with CTC loss

HMM

LSTM on OCR

• no explicit restriction on the amount of context they can learn
• shown to learn character language models (implicit LM)
• Implicit LM helps improve performance on the OCR task

Context

Experiment

Shuffled Characters Experiment

1. 找一篇文章當對照組

2. 把文章的字打亂當實驗組

hypothesis

Implicit LM 會幫助對照組達到更低的 CER (character error rate)

Shuffled Characters Experiment

N-gram Experiment

找到 implicit LM 的極限在哪

覺得這個實驗是湊數用的

 ... the performance will improve as length increases and plateau where implicit LM stops considering more context frames.

N-gram Experiment

• Seen 出現過的 n-gram
• Unseen 沒出現過的 n-gram 
• Purely Unseen 所有 substring 都是沒出現過的 n-gram

N-gram Experiment