EasyOCR-onnx / english_g2_jpqd.yaml
asmud's picture
Initial release: EasyOCR ONNX models with JPQD quantization
c1ac2fb
name: english_g2_jpqd
description: English text recognition model (CRNN) optimized with JPQD quantization
framework: ONNX
task: text-recognition
domain: computer-vision
subdomain: optical-character-recognition
model_info:
architecture: CRNN (CNN + BiLSTM + CTC)
language: English
character_set: "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~"
num_classes: 95
original_source: EasyOCR
optimization: JPQD quantization
specifications:
input_shape: [1, 1, 32, 100]
input_type: float32
input_format: Grayscale
output_shape: [1, 25, 95] # sequence_length x num_classes
output_type: float32
batch_size: dynamic
sequence_length: 25
performance:
original_size_mb: 14.4
optimized_size_mb: 8.5
compression_ratio: 3.97
inference_time_cpu_ms: ~10
accuracy_retention: ">95%"
deployment:
runtime: onnxruntime
hardware: CPU-optimized
precision: INT8 weights, FP32 activations
memory_usage_mb: ~15
usage:
preprocessing:
- Convert to grayscale
- Resize to 32x100 (height x width)
- Normalize to [0,1]
- Add batch and channel dimensions
postprocessing:
- Apply CTC decoding
- Convert indices to characters
- Remove blank tokens and duplicates
supported_characters:
digits: "0-9"
lowercase: "a-z"
uppercase: "A-Z"
punctuation: "!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~"
training_data:
type: Synthetic and real text images
languages: English
domains: Documents, natural scenes, printed text
license: apache-2.0
tags:
- text-recognition
- english
- crnn
- lstm
- ocr
- onnx
- quantized
- jpqd