en_PP-OCRv5_mobile_rec

Recognizes text within pre-cropped textline image regions using a lightweight CNN-RNN architecture optimized for mobile deployment. The model processes variable-length textline images through a ResNet backbone for feature extraction, followed by a bidirectional LSTM sequence decoder that outputs character-level predictions. Architecture uses attention mechanisms to handle variable text lengths and orientations, with quantization and pruning applied to reduce model size from ~200MB to ~8-10MB for on-device inference.

Decodes variable-length character sequences from textline feature maps using a bidirectional LSTM with attention mechanism. The decoder attends over spatial feature dimensions to predict characters sequentially, handling text of different lengths (typically 1-50 characters) without fixed-size constraints. Attention weights allow the model to focus on relevant image regions for each predicted character, improving accuracy on compressed or distorted text.

Extracts spatial feature representations from textline images using a lightweight ResNet backbone (typically ResNet18 or ResNet34 variant) with depthwise separable convolutions for mobile efficiency. The backbone progressively downsamples spatial dimensions while increasing channel depth, producing feature maps that capture character-level visual patterns (strokes, curves, spacing). Intermediate feature maps are concatenated to preserve multi-scale information critical for recognizing text at different scales and resolutions.

Deploys the recognition model on mobile devices using INT8 quantization and PaddleLite runtime, reducing model size from ~200MB (FP32) to ~8-10MB (INT8) with minimal accuracy loss (<1%). Quantization is applied post-training using calibration data; the model is converted to PaddleLite format with operator fusion and memory layout optimization for ARM processors. Inference runs directly on mobile CPUs without GPU dependency, achieving 10-50ms latency per textline on modern mobile hardware.

Preprocesses variable-width textline images into normalized batches for inference, handling resizing, padding, and channel normalization. Images are resized to fixed height (32 pixels) while preserving aspect ratio, padded to a common width within the batch, and normalized using ImageNet statistics (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]). Preprocessing is implemented in C++ for PaddleLite and Python for server inference, with SIMD optimizations for mobile platforms.

Extracts character-level confidence scores from model output logits and applies post-processing filters to remove low-confidence predictions. The model outputs logits for each character position; softmax is applied to convert to probabilities, and per-character confidence is extracted as the maximum probability. Filtering strategies include: removing characters with confidence <threshold, merging adjacent low-confidence predictions, and flagging uncertain regions for manual review. Confidence scores enable downstream applications to prioritize high-confidence text for processing.

Designed as the recognition stage of PaddleOCR's two-stage pipeline, consuming textline bounding boxes and cropped images from the detection model (en_PP-OCRv5_mobile_det). The recognition model expects pre-cropped textline images with minimal padding; integration requires coordinate transformation from detection output (rotated bounding boxes) to axis-aligned crops. PaddleOCR provides end-to-end orchestration via the OCRv5 inference API, handling detection→crop→recognition→post-processing in a single call.