trocr-base-printed

Converts images of printed text documents into machine-readable text using a vision-encoder-decoder transformer architecture. The model encodes visual features from document images through a CNN-based vision encoder, then decodes those features into character sequences using an autoregressive text decoder. Specifically optimized for printed (non-handwritten) documents with clear typography, handling multi-line text recognition through sequential token generation with attention mechanisms over spatial image regions.

Automatically preprocesses document images to optimal input specifications for the vision encoder, including resizing to 384x384 pixels, channel normalization using ImageNet statistics, and padding/cropping to maintain aspect ratios. Handles variable input image sizes and formats through a standardized pipeline that converts raw images into normalized tensor batches compatible with the encoder's expected input shape and value ranges.

Generates text output character-by-character using an autoregressive transformer decoder that conditions on previously generated tokens and visual encoder features. Implements beam search decoding (with configurable beam width, typically 4-8) to explore multiple hypothesis sequences in parallel, selecting the highest-probability complete sequence rather than greedy single-token selection. This enables recovery from early decoding errors and improves overall text accuracy through probabilistic search over the output space.

Extracts spatial attention weights from the decoder's cross-attention mechanism over encoder features, enabling identification of which image regions correspond to generated text tokens. The decoder produces attention maps (shape [num_heads, seq_len, spatial_positions]) that indicate which parts of the input image were attended to when generating each output character. These attention weights can be visualized as heatmaps or used to extract bounding boxes for individual words or characters within the document image.

Recognizes printed text in multiple languages using a language-agnostic vision encoder trained on diverse scripts and character sets. The encoder learns visual features that generalize across Latin, Cyrillic, Arabic, CJK, and other scripts without language-specific preprocessing. The decoder is trained on multilingual text corpora, enabling character-level generation across supported languages. Language identification is implicit through the decoder's learned character distributions rather than explicit language tags.

Supports deployment optimization through quantization (INT8, FP16) and compatibility with distilled model variants that reduce model size and inference latency. The base model can be quantized post-training using standard PyTorch/TensorFlow quantization tools, reducing model size from ~347MB to ~87MB (INT8) with minimal accuracy loss. Distilled variants (trocr-small-printed) are available as pre-trained checkpoints, offering 3-4x faster inference with ~2-3% accuracy degradation for resource-constrained deployments.