界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

dbnet

dbnet

【效率与协作】Use EasyOCR's DBNet detector, check DBNet import/init, and compile the DCN operator when the detector backend needs repair.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: dbnet description: "Use EasyOCR's DBNet detector, check DBNet import/init, and compile the DCN operator when the detector backend needs repair." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

EasyOCR DBNet

Use this sub-skill when the task involves detect_network='dbnet18', DBNet import/init, or compiling the DCN operator that DBNet needs.

Use when

  • The user wants the DBNet detector instead of the default CRAFT detector.
  • The user asks about DBNet, dcn, compile, NVCC, or CUDA/CPU build issues.
  • A DBNet import works but initialization or detection fails.
  • The user needs a quick CPU smoke for the DBNet module.

Do not use when

  • The task is ordinary OCR with the default detector. Use sub-skills/inference/.
  • The task is about custom recognition bundles. Use sub-skills/custom-models/.
  • The task is only about training or ONNX export.

Quick workflow

  1. Read references/workflows.md for the DBNet runtime and compile sequence.
  2. Run scripts/compile_dcn.py --help or --check-only to inspect the current DCN state.
  3. If needed, run scripts/compile_dcn.py --build after checking the compiler prerequisites.
  4. Use references/troubleshooting.md for device mismatch, missing operator, and compiler failures.

What this sub-skill owns

  • DBNet detector selection inside EasyOCR.
  • DBNet import/init checks.
  • DCN operator build and presence checks.
  • CPU versus CUDA compilation guidance for the DBNet operator.

References and scripts

  • ../../references/api-reference.md for the main Reader surface.
  • ../../references/configuration.md for cache locations and detector choice.
  • references/workflows.md for DBNet runtime and compile steps.
  • references/troubleshooting.md for backend, operator, and device errors.
  • scripts/compile_dcn.py for a bundled DCN helper.

Boundary notes

  • Keep the default detector workflow in inference.
  • Keep custom model loading in the custom-models sub-skill.
  • Treat dbnet50 as an internal DBNet detail here; Reader only exposes the DBNet choice used in this checkout.
我的收藏