For Chinese: 详细做法请参考我的博客

CRNN

A pytorch implementation of CRNN，and test it with IIIT-5K.
Support PyTorch 1.0 now.

Paper.

Netword Struct：

Type	Configurations	Output Size
Input	W × 32 gray-scale image	W × 32 × 1
Convolution	#maps:64, k:3 × 3, s:1, p:1	W × 32 × 64
MaxPooling	Window:2 × 2, s:2	W/2 × 16 × 64
Convolution	#maps:128, k:3 × 3, s:1, p:1	W/2 × 16 × 128
MaxPooling	Window:2 × 2, s:2	W/4 × 8 × 128
Convolution	#maps:256, k:3 × 3, s:1, p:1	W/4 × 8 × 256
Convolution	#maps:256, k:3 × 3, s:1, p:1	W/4 × 8 × 256
MaxPooling	Window:1 × 2, s:2	W/4 × 4 × 256
Convolution	#maps:512, k:3 × 3, s:1, p:1	W/4 × 4 × 512
BatchNormalization	-	W/4 × 4 × 512
Convolution	#maps:512, k:3 × 3, s:1, p:1	W/4 × 4 × 512
BatchNormalization	-	W/4 × 4 × 512
MaxPooling	Window:1 × 2, s:2	W/4 × 2 × 512
Convolution	#maps:512, k:2 × 2, s:1, p:0	W/4-1 × 1 × 512
Map-to-Sequence	-	W/4-1 × 512
Bidirectional-LSTM	#hidden units:256	W/4-1 × 256
Bidirectional-LSTM	#hidden units:256	W/4-1 × label_num
Transcription	-	str

Requirements

• Python 3
• PyTorch 1.0

Usage

Click here and download IIIT-5K dataset to the 'data/' folder of current path.

python main.py --fix_depth

You can tune the hyperparameters for better performence.