clovaai/deep-text-recognition-benchmark

Features

• Optical Character Recognition - Provides a full system for converting images of text in natural scenes into machine-encoded text.
• Connectionist Temporal Classification - Utilizes connectionist temporal classification to map variable-length predicted sequences to target text by collapsing repeats.
• Scene Text Recognition Frameworks - Provides a PyTorch-based deep learning framework for extracting text from images using visual and sequential stages.
• Deep Learning Pipelines - Implements a full computational workflow from training data preparation through model training and validation.
• Hybrid CNN-RNN Architectures - Combines convolutional layers with recurrent networks and CTC for end-to-end sequence-based text recognition.
• Model Performance Benchmarking - Provides standardized tests to evaluate recognition accuracy, inference speed, and memory usage across consistent datasets.
• OCR Model Training - Provides specialized training processes to teach neural networks to recognize visual text and varying fonts.
• Scene Text Recognition - Identifies and reads characters from complex real-world images with varying backgrounds and perspectives.
• OCR Benchmarks - Provides a standardized system for measuring the accuracy and inference speed of text recognition models.
• Sequence-to-Sequence Mappings - Processes visual features through a recurrent neural network to maintain spatial order and character dependencies.
• Optical Character Recognition Training - Provides a complete pipeline for training deep learning models to recognize text from scene images.
• Text Recognition Benchmarks - Measures and compares the accuracy and inference speed of different text recognition models using standardized datasets.
• Convolutional Feature Extraction - Extracts high-level visual spatial patterns from images using deep convolutional layers before sequence decoding.
• Inference Latency Measurement - Measures average GPU processing time per image to determine the inference efficiency of the recognition model.
• Recognition Accuracy Evaluation - Provides tools for quantifying recognition quality using precision and normalized edit distance across benchmark datasets.
• Index-to-Text Conversions - Translates numerical model outputs into human-readable text by removing repeated characters and blank tokens.
• Text-to-Numeric Transformations - Transforms raw text labels into numerical indices while handling padding and special tokens for loss functions.
• Training Dataset Balancing - Implements techniques to adjust the sampling frequency of multiple datasets within training batches to prevent model bias.
• Training Dataset Preparation - Transforms raw images and text labels into optimized formats for efficient deep learning model training.
• Edit Distance Calculators - Quantifies recognition accuracy by calculating the minimum number of operations to transform predictions into ground truth.
• Text Recognition Models - Comprehensive benchmark and comparison framework for text recognition models.
• Text recognition - Listed in the “Text recognition” section of the Ailia Models awesome list.

Star history