ArmSoM
← All stories

Local Document OCR on an ARM SBC: Model and Memory Planning

Local OCR combines image preparation, text detection and recognition. A board may accelerate one stage while the rest stays on the CPU. Evaluate the entire document task, especially when inputs include rotation, small text or uneven lighting.

ArmSoM Sige7

ArmSoM Sige7

Build a document set before tuning

Include the formats, languages and image quality the application will receive. Record recognition errors as well as processing time. Check image scaling, orientation and text-region handling before changing the model. If an accelerator is used, identify which OCR stages it executes and measure the remaining CPU work separately.

Profile each stage

On Sige5 or Sige7, measure image decoding and preprocessing as well as model execution. Large pages create sizable intermediate buffers, and processing several pages concurrently can multiply memory use. Check that the chosen model components have a supported deployment path before assuming the NPU accelerates the whole pipeline.

Design output and failure handling

Return confidence or a review state where appropriate. Preserve the original document and page mapping so an operator can inspect uncertain results. Limit job size and concurrency to avoid an oversized upload disrupting other services. Compare extracted fields against a ground-truth sample before describing the system as suitable for unattended document entry.

Match the choice to your requirements

Requirement or situation

What to evaluate

Decision detail

Input

Page size, language and layout

Test difficult examples.

Runtime

Stage timing and peak memory

Identify the real bottleneck.

Output

Field accuracy and review path

Keep traceability to the original page.

Example design

Build an OCR acceptance set around the documents

An invoice, a photograph of a sign and a scanned manual present different problems. Collect representative layouts, lighting, language and font sizes, including difficult examples. Evaluate detection, recognition and reading order separately. Measure time per complete document rather than only the neural-network step. If the application extracts structured fields, verify those fields directly; a readable text preview can still contain an incorrect amount or identifier.

Hardware and documentation

Sige7 · RK3588

RK3588, 6 TOPS NPU, dual 2.5GbE and M.2 PCIe 3.0 x4. Check the selected eMMC and wireless configuration.

Hardware documentation · Product details

Sige5 · RK3576

RK3576, four Cortex-A72 and four Cortex-A53 cores, 6 TOPS NPU, dual Gigabit Ethernet and M.2 PCIe 2.1 x1.

Hardware documentation · Product details

Frequently asked questions

Does OCR support mean every language works equally well?

No. Model training, text size, layout and image quality determine results.

Product information and support

Explore Sige7

Contact ArmSoM

Sources

Written by ArmSoM using the documentation below. Example designs are proposals, not measured customer installations.

Related reading