PP-OCRv6-tiny optical character recognition compiled for the Synaptics Torq NPU:
DBNet text detection followed by CTC text recognition. Both stages run on the
NPU. Used by the ppocr demo in
torq-examples.
The recognition dictionary is Chinese + English (6,904 characters), so Latin
text, digits and punctuation decode natively; Japanese, Korean, Cyrillic and
Arabic are not covered.
Input — samples/sample.jpg
NPU output — boxes + recognized text
All ten lines are read correctly at confidence ≥ 0.966, in 1.7 s end to end on
an SL2619.
Files
File
Purpose
ppocr_det_800x608.vmfb
Detection (DBNet), static 800×608 bf16 input
ppocr_det_640x352.vmfb
Detection (DBNet), static 640×352 bf16 input, for wide/16:9 sources
rec_buckets/rec_w320.vmfb
Recognition, 48×320 lines
rec_buckets/rec_w640.vmfb
Recognition, 48×640 lines
rec_buckets/rec_w1280.vmfb
Recognition, 48×1280 lines
rec_buckets/rec_w2432.vmfb
Recognition, 48×2432 lines
ppocr_rec.yml
Recognizer character dictionary
ppocr_det_dynamic.onnx
fp32 detection, CPU reference for accuracy checks
ppocr_rec_dynamic.onnx
fp32 recognition, CPU reference for accuracy checks
samples/sample.jpg
Sample café menu card, 10 text lines
Two detector sizes
DBNet's stride-32 backbone requires input dims that are multiples of 32.
800×608 is the default, matched to portrait documents; 640×352 serves wide/16:9
sources — it is 640×360 rounded up to the next multiple of 32, so letterboxed
16:9 content is padded rather than cropped. Select it in the demo with
--det-hw 640 352.
Why four recognition models
Recognition input width is static per vmfb. Each detected line is routed to the
narrowest bucket it fits in, so a short label is padded to 320 rather than to
the widest width. Lines longer than 2432 clamp to the widest bucket.
Either stage can be switched to ONNX Runtime with --det-backend ort /
--rec-backend ort (plus the matching --det-onnx / --rec-onnx) to compare
NPU output against a CPU reference.
Measured on SL2619
samples/sample.jpg, 912×1200, 10 text lines detected:
Stage
Time
Detection (800×608)
~0.53 s
Recognition (10 lines, bucketed)
~1.19 s
Recognition scales with the number of detected lines, because each line is a
separate invocation — the bucket models are compiled with a static batch of 1.
A dense page of 99 lines takes roughly 22 s.
About sample.jpg
A 912×1200 café menu card, rendered synthetically in DejaVu Serif rather than
photographed, so it carries no third-party image licensing.
Its width/height ratio of 0.76 matches the detector's static 608×800 input.
Preprocessing resizes straight to that shape without preserving aspect, so an
off-ratio image reaches the model stretched — worth matching if you swap in your
own sample.
Recognized output, all ten lines at confidence ≥ 0.966:
1 [0.991] BLUE DOOR CAFE 6 [0.995] Smoked Salmon Bagel 9.75
2 [0.996] all day breakfast 7 [1.000] DRINKS
3 [0.999] BREAKFAST 8 [0.975] Espresso 2.75
4 [0.996] Avocado Toast 6.50 9 [0.966] Fresh Orange Juice 4.00
5 [0.996] Buttermilk Pancakes 7.00 10 [0.993] open 7am - 3pm daily
