ML Tasks Image OCR SW

[junxnone]

Sliding Windows CNN CTC

Reference

• paper - 2017 - Scene Text Recognition with Sliding Convolutional Character Models
• 场景文字识别Sliding Convolution笔记
• Sequence Modeling With CTC

Brief

• Sliding windows + CNN + CTC
• English & Chinese
  • Softmax Layer - 37/7357 English/Chinese

Arch

Name	Description
Sliding window	crop bbox x N
Classification Layer	feature --> label
CTC Transcription layer	label --> final predict result

Preprocess & Sliding Windows

Preprocess	Training	Testing
Input scale to Height = 32, Width 按比例缩放 (Pad -> 256)	✅	✅
Window size	- Single-scale WxH = 32x32 - Multi-scale WxH = 32x24 - WxH = 32x32 - WxH = 32x40 - > resize to 32x32	32x32
Step Size = 4	✅	✅

Pipeline	CNN Layers	CTC Algo

sliding windows examples

decoding

• naive - None - lexicon -free
• lexicon based
• language model based

[junxnone]

junxnone/tech-io#749