Textline Orientation Classification Via Lightweight Cnn

image-to-text model by undefined. 2,05,933 downloads.

Unique: PP-LCNet architecture uses depthwise-separable convolutions with SE (squeeze-and-excitation) blocks to achieve <2MB model size while maintaining competitive accuracy on textline orientation — specifically designed for the PaddleOCR pipeline rather than generic image classification, enabling tight integration with text detection and recognition stages.

vs others: Smaller and faster than general-purpose image classifiers (ResNet, EfficientNet) for this specific task, with native PaddleOCR integration eliminating format conversion overhead; outperforms rule-based angle detection on degraded documents.

via “mobile-optimized textline recognition from image crops”

image-to-text model by undefined. 3,39,341 downloads.

Unique: Uses PaddleOCR's proprietary lightweight architecture combining ResNet feature extraction with bidirectional LSTM decoding, specifically tuned for mobile inference via PaddleLite quantization (INT8/FP16). Unlike generic CRNN models, incorporates attention mechanisms for variable-length handling and applies knowledge distillation to reduce parameters by ~60% while maintaining accuracy parity with full models.

vs others: Smaller model footprint (~8-10MB) than Tesseract or EasyOCR with faster mobile inference, and better accuracy on modern fonts than traditional Tesseract; trades off language diversity for English-specific optimization and requires detection model pairing.