OCR with CRNN in PyTorch and Python 3: Handling Variable-Length Chinese Characters

OCR (Optical Character Recognition) with CRNN (Convolutional Recurrent Neural Network) is a popular approach for recognizing variable-length text in images. In this tutorial, we’ll focus on implementing OCR for Chinese characters using PyTorch and Python 3. We’ll cover the essential steps from data preprocessing to model training and evaluation.
1. Data Preprocessing
Data preprocessing is crucial for accurate OCR. Common steps include image resize, normalization, and converting the image to a tensor.
Here’s an example of preprocessing using Python and OpenCV:

import cv2
import numpy as np
import torch
def preprocess_image(image_path):
# Read the image
image = cv2.imread(image_path)
# Resize the image to a fixed size (e.g., 32x100)
resized_image = cv2.resize(image, (100, 32))
# Normalize the image (subtract mean and divide by standard deviation)
mean = np.array([127.5])  # mean for grayscale images
std = np.array([127.5])  # standard deviation for grayscale images
normalized_image = (resized_image - mean) / std
# Convert the image to a PyTorch tensor
tensor_image = torch.from_numpy(normalized_image).unsqueeze(0).float()
return tensor_image

2. Model Architecture
CRNN consists of several components: convolutional layers for feature extraction, recurrent layers (LSTM or GRU) for capturing temporal dependencies, and a fully connected layer for classification.
Here’s a simplified example of the CRNN architecture in PyTorch:
```python
import torch.nn as nn
class CRNN(nn.Module):
def init(self, numclasses=37): # Assuming 37 Chinese characters plusblank
super(CRNN, self)._init()
self.conv = nn.Sequential(
nn.Conv2d(1, 64, kernel_size=(3, 3), padding=1),
n.ReLU(),
n.MaxPool2d(kernel_size=2, stride=2),
n.Conv2d(64, 128, kernel_size=(3, 3), padding=1),
n.ReLU(),
n.MaxPool2d(kernel_size=2, stride=2),
n.Conv2d(128, 256, kernel_size=(3, 3), padding=1),
n.ReLU(),
n.MaxPool2d(kernel_size=2, stride=2),
n.Conv2d(256, 256, kernel_size=(3, 3), padding=1),
n.ReLU(),
n.MaxPool2d(kernel_size=2, stride=2),
n.Conv2d(256, 512, kernel_size=(3, 3), padding=1),
n.BatchNorm2d(512),
n.ReLU(),
n.MaxPool2d(kernel_size=(2, 1)),
n.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
n.BatchNorm2d(512),
n.ReLU(),
n.MaxPool2d(kernel_size=(1, 2)),
n.Conv2d(512, 512, kernel_size=(3, 3), padding=1),
n.BatchNorm2d(512),
n.ReLU(),
n.MaxPool2d(kernel_size=(1, 1)),
n.
)
self.rnn = nn.Sequential(
nLSTM(512, 256),
nLinear