Fruit Inspector

Build 0.10.0 · Windows x64 · Linux x64 · Nim · ONNX Runtime (CPU) · conveyor station demo (eval)

Fruit Inspector is conveyor inspection: fruit on a belt; quality is graded only in the control zone (centre cell C), not anywhere in the frame.

• This repo: runnable Windows and Linux demos (ImGui shell + inference libraries). No Nim product source on main.
• Same three inference DLLs for your own UI: fruit-inspector-engine · HOST.md

Download

File	Use
Fruit_Inspector_DLL.7z	Windows (0.10.0): Fruit_Inspector.exe, three DLLs, ORT, apple models. Use this on Win64.
Fruit_Inspector_linux_x64.zip	Linux x64 (0.10.0): portable folder — binary, three .so, bundled ORT, models, run.sh. Debian 12/13, Ubuntu 22.04+.
Fruit_Inspector.7z	Older Windows archive on the same release; not the full three-DLL layout. Reference only.

All assets: Release 0.10.0

Windows: unpack with 7-Zip or similar. Linux: unzip the zip. Use a local folder, not a network drive.

---

What is on main

• README.md, LICENSE.md, licenses/
• Demo media: infer.gif, processor.jpg, speed.jpg
• Binaries live in Releases only

Contents of Fruit_Inspector_DLL.7z

• Fruit_Inspector.exe — UI, 3×3 grid, conveyor focus/sticky, Windows camera (NV12)
• fruit_infer.dll, fruit_cat.dll, fruit_features.dll — all ONNX Runtime inference; the exe is a shell (no in-process ORT)
• lib/cpu/onnxruntime.dll
• model/apple/ — det.onnx, cat.onnx, features.json (apple pack only)
• settings.json, licenses/ (keep the whole folder)

Not in the default archive:

• lib/openvino/, lib/cuda/ — bring your own if you switch EP; default is "ort_provider": "cpu"
• trig.exe — optional LAN helper (see trig); usually not bundled
• Pan-tilt / neck — off in this drop (neck_com_port: 0)

Contents of Fruit_Inspector_linux_x64.zip

• Fruit_Inspector — main app (ext4 / home disk); same binary as Fruit_Inspector.exe for FAT USB launchers
• fruit_infer.so, fruit_cat.so, fruit_features.so — ONNX Runtime inference; the app is a shell
• lib/linux/ — bundled ONNX Runtime CPU
• model/apple/, settings.json, licenses/
• run.sh, install_to_home.sh, Fruit_Inspector.desktop — launch (use run.sh on USB; see README.txt inside the zip)

CPU: x86_64 with SSE4.2 + AVX1 (AVX2 not required for this demo drop). Not ARM / Raspberry Pi.

Not offered here yet:

• ARM demo
• Sample host app (C#, Lazarus, …) — see the engine repo
• Factory PLC chain (stop belt → grade) — direction only, not shipped
• Other crops — apple only unless added later

---

Pictures

Test bench: Intel i5-11400, 32 GB RAM, no discrete GPU, USB camera 1920×1080.

Speed (that bench, CPU EP): With dual YOLO (640 full frame + 256 ROI refine) and the full loop (preprocess + ORT + postprocess), the console infer loop is about 115 fps — lower than with only the main detect path; the second model and extra pre/post cost time. Live camera preview is about 30 fps (camera limit, not infer). speed.jpg may show an older setup; trust [infer] when you run from cmd.

---

Idea

• 3×3 grid; C = control zone
• Up to two apples on the line; stable IDs; focus on nearest to C; sticky after leave C
• Cat (LightGBM) only when the focus apple centre is inside C
• Pipeline: YOLOX 640 → ROI 256 → 44 features → Cat → GOOD / BAD on screen

---

Station behaviour

• No full-frame chase — work in C with an ID queue
• Approaching C: detect and track only
• In C: features and verdict
• Left C: sticky handoff to the next ID

---

Architecture (0.10.0)

Camera (NV12) → Fruit_Inspector (UI + conveyor)
                      ↓
              fruit_infer (.dll / .so) (YOLO)
                      ↓
              RGB crop on host
                      ↓
              fruit_features → fruit_cat

---

trig

Optional LAN client for GOOD/BAD over TCP from the inspector. Full line story (stop section → ready → grade) is not implemented; on the desk, Cat runs on enter C.

---

Stack

• Nim, ImGui (GLFW); camera: Windows Media Foundation or Linux V4L2 → NV12
• ONNX Runtime CPU in the release archives
• settings.json: provider, models, camera, threads — edit by hand or Settings → Apply

---

Run (Windows)

1. Download Fruit_Inspector_DLL.7z and unpack.
2. One tree: exe, three DLLs, lib/cpu/, model/apple/, settings.json, licenses/.
3. Set camera_index and ort_provider (cpu first).
4. From cmd: Fruit_Inspector.exe for [infer] fps on stdout.
5. UI: START / STOP; Esc confirms quit.

Run (Linux)

1. Download Fruit_Inspector_linux_x64.zip and unzip.
2. chmod +x run.sh (ext4) or always: bash /path/to/Fruit_Inspector_linux_x64/run.sh.
3. Install once: GL/Mesa/X11 libs, v4l-utils; user in group video for USB camera.
4. On FAT USB: do not run bare Fruit_Inspector — use .desktop, Fruit_Inspector.exe, or run.sh (details in zip README.txt).
5. Same UI behaviour as Windows; logs under logs/.

---

License

Testing, teaching, research, and feedback — yes. Commercial production under this license — no.

LICENSE.md, licenses/Fruit_Inspector_LICENSE.md, licenses/THIRD_PARTY.md. Keep licenses/ next to the exe.

Commercial inquiries: olesha-ai.

Intel, OpenVINO, Microsoft, Megvii, and other names are trademarks of their respective owners.

---

olesha-ai