TrOCR: Optimized for Qualcomm Devices

End-to-end text recognition approach with pre-trained image transformer and text transformer models for both image understanding and wordpiece-level text generation.

This is based on the implementation of TrOCR found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.45, ONNX Runtime 1.25.0	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit TrOCR on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

Custom weights (e.g., fine-tuned checkpoints)
Custom input shapes
Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for TrOCR on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_to_text

Model Stats:

Model checkpoint: trocr-small-stage1
Input resolution: 320x320
Number of parameters (decoder): 38.3M
Model size (decoder) (float): 146 MB
Number of parameters (encoder): 23.0M
Model size (encoder) (float): 87.8 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
decoder	ONNX	float	Snapdragon® X2 Elite	1.259 ms	7 - 7 MB	NPU
decoder	ONNX	float	Snapdragon® X Elite	2.1 ms	68 - 68 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	1.449 ms	0 - 282 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Gen 1 Mobile	2.735 ms	1 - 218 MB	NPU
decoder	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	2.112 ms	1 - 3 MB	NPU
decoder	ONNX	float	Qualcomm® QCS8450	2.735 ms	1 - 218 MB	NPU
decoder	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	2.74 ms	7 - 16 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	1.182 ms	2 - 179 MB	NPU
decoder	ONNX	float	Snapdragon® 8 Elite Mobile	1.3 ms	0 - 189 MB	NPU
decoder	ONNX	float	Qualcomm® Dragonwing™ Q-8750	1.3 ms	0 - 189 MB	NPU
decoder	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	2.1 ms	68 - 68 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X2 Elite	1.556 ms	7 - 7 MB	NPU
decoder	QNN_DLC	float	Snapdragon® X Elite	2.208 ms	7 - 7 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	1.424 ms	0 - 278 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	2.672 ms	2 - 216 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8275	4.137 ms	5 - 106 MB	NPU
decoder	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	2.012 ms	1 - 2 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8775P	2.861 ms	5 - 107 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8650P	2.861 ms	5 - 107 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8255P	2.861 ms	5 - 107 MB	NPU
decoder	QNN_DLC	float	Qualcomm® QCS8450	2.672 ms	2 - 216 MB	NPU
decoder	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	2.859 ms	7 - 16 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	1.158 ms	1 - 172 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA7255P	4.137 ms	5 - 106 MB	NPU
decoder	QNN_DLC	float	Snapdragon® 8 Elite Mobile	1.247 ms	0 - 187 MB	NPU
decoder	QNN_DLC	float	Qualcomm® SA8295P	2.558 ms	0 - 42 MB	NPU
decoder	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	1.247 ms	0 - 187 MB	NPU
decoder	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	2.208 ms	7 - 7 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	1.381 ms	0 - 280 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	2.456 ms	0 - 217 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS8275	12.741 ms	7 - 22 MB	GPU
decoder	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	1.95 ms	0 - 2 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8775P	2.847 ms	0 - 103 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8650P	2.847 ms	0 - 103 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8255P	2.847 ms	0 - 103 MB	NPU
decoder	TFLITE	float	Qualcomm® QCS8450	2.456 ms	0 - 217 MB	NPU
decoder	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	2.779 ms	0 - 83 MB	NPU
decoder	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	1.136 ms	0 - 171 MB	NPU
decoder	TFLITE	float	Qualcomm® SA7255P	12.741 ms	7 - 22 MB	GPU
decoder	TFLITE	float	Snapdragon® 8 Elite Mobile	1.266 ms	0 - 190 MB	NPU
decoder	TFLITE	float	Qualcomm® SA8295P	2.597 ms	0 - 44 MB	NPU
decoder	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	1.266 ms	0 - 190 MB	NPU
encoder	ONNX	float	Snapdragon® X2 Elite	4.557 ms	16 - 16 MB	NPU
encoder	ONNX	float	Snapdragon® X Elite	11.3 ms	48 - 48 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Gen 3 Mobile	7.834 ms	0 - 199 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Gen 1 Mobile	19.787 ms	0 - 315 MB	NPU
encoder	ONNX	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	10.93 ms	0 - 269 MB	NPU
encoder	ONNX	float	Qualcomm® QCS8450	19.787 ms	0 - 315 MB	NPU
encoder	ONNX	float	Qualcomm® Dragonwing™ IQ-9075	13.736 ms	15 - 19 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	4.149 ms	0 - 124 MB	NPU
encoder	ONNX	float	Snapdragon® 8 Elite Mobile	5.304 ms	16 - 139 MB	NPU
encoder	ONNX	float	Qualcomm® Dragonwing™ Q-8750	5.304 ms	16 - 139 MB	NPU
encoder	ONNX	float	Qualcomm® Dragonwing™ IQ-X7181	11.3 ms	48 - 48 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X2 Elite	5.137 ms	2 - 2 MB	NPU
encoder	QNN_DLC	float	Snapdragon® X Elite	12.101 ms	2 - 2 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	8.005 ms	2 - 192 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Gen 1 Mobile	19.346 ms	2 - 309 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8275	37.051 ms	2 - 107 MB	NPU
encoder	QNN_DLC	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	11.262 ms	2 - 4 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8775P	14.018 ms	2 - 109 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8650P	14.018 ms	2 - 109 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8255P	14.018 ms	2 - 109 MB	NPU
encoder	QNN_DLC	float	Qualcomm® QCS8450	19.346 ms	2 - 309 MB	NPU
encoder	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-9075	14.22 ms	4 - 14 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	4.351 ms	2 - 115 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA7255P	37.051 ms	2 - 107 MB	NPU
encoder	QNN_DLC	float	Snapdragon® 8 Elite Mobile	5.44 ms	2 - 116 MB	NPU
encoder	QNN_DLC	float	Qualcomm® SA8295P	19.681 ms	2 - 235 MB	NPU
encoder	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	5.44 ms	2 - 116 MB	NPU
encoder	QNN_DLC	float	Qualcomm® Dragonwing™ IQ-X7181	12.101 ms	2 - 2 MB	NPU
encoder	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	7.705 ms	3 - 195 MB	NPU
encoder	TFLITE	float	Snapdragon® 8 Gen 1 Mobile	19.56 ms	7 - 312 MB	NPU
encoder	TFLITE	float	Qualcomm® QCS8275	36.827 ms	7 - 119 MB	NPU
encoder	TFLITE	float	Qualcomm® Dragonwing™ QCS8550 (Proxy)	10.673 ms	7 - 9 MB	NPU
encoder	TFLITE	float	Qualcomm® SA8775P	13.49 ms	7 - 111 MB	NPU
encoder	TFLITE	float	Qualcomm® SA8650P	13.49 ms	7 - 111 MB	NPU
encoder	TFLITE	float	Qualcomm® SA8255P	13.49 ms	7 - 111 MB	NPU
encoder	TFLITE	float	Qualcomm® QCS8450	19.56 ms	7 - 312 MB	NPU
encoder	TFLITE	float	Qualcomm® Dragonwing™ IQ-9075	13.235 ms	6 - 65 MB	NPU
encoder	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	4.119 ms	0 - 104 MB	NPU
encoder	TFLITE	float	Qualcomm® SA7255P	36.827 ms	7 - 119 MB	NPU
encoder	TFLITE	float	Snapdragon® 8 Elite Mobile	5.202 ms	6 - 119 MB	NPU
encoder	TFLITE	float	Qualcomm® SA8295P	19.106 ms	7 - 237 MB	NPU
encoder	TFLITE	float	Qualcomm® Dragonwing™ Q-8750	5.202 ms	6 - 119 MB	NPU

License

The license for the original implementation of TrOCR can be found here.

References

Community

Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
For questions or feedback please reach out to us.

Downloads last month: -; Downloads are not tracked for this model. How to track

Space using qualcomm/TrOCR 1

Paper for qualcomm/TrOCR

TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models

Paper • 2109.10282 • Published Sep 21, 2021 • 13