References

Vendor documents, standards and papers the design rests on. Cited in the text as [n].

Vendor and standards

[1] BrainChip Inc. Akida / MetaTF documentation, version 2.19.x. doc.brainchipinc.com. — Akida 1.0 layer types (InputConvolutional, Convolutional, SeparableConvolutional, FullyConnected), on-chip unsupervised learning, and the note that the power meter is currently supported on the AKD1000 only.

[2] Brainchip-Inc. akida_dw_edma — PCIe driver for AKD1500, GitHub repository. — Official baseline for guide chapter 05; our two kernel-6.12 patches are not part of it.

[3] ISA-101.01-2015. Human Machine Interfaces for Process Automation Systems. International Society of Automation. — Basis of the UI 2.2 principles: grey surfaces, colour only for state, alarm philosophy, operator/expert separation.

[4] Raspberry Pi Ltd. Camera documentation and libcamera controls (Camera Module 3 / IMX708). — AfMode, AeEnable, ExposureTime, AnalogueGain, ColourGains used by the camera strip.

Detection and assignment (SE-M3)

[5] Bodla, N., Singh, B., Chellappa, R., Davis, L. S. Soft-NMS — Improving Object Detection with One Line of Code. ICCV 2017. — The slot-aware suppression between grid cells in SE-M3 is a Soft-NMS on rotated boxes; suppression is never applied between slots of the same cell.

[6] Lin, T.-Y., Goyal, P., Girshick, R., He, K., Dollár, P. Focal Loss for Dense Object Detection. ICCV 2017. — Focal weighting of objectness against the ~1:500 slot imbalance in the 32 × 32 grid.

[7] Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S. End-to-End Object Detection with Transformers. ECCV 2020. — Set prediction with bipartite matching; SE-M3 applies the same idea locally with an exhaustive permutation search (M! ≤ 24) instead of the Hungarian algorithm.

[8] Redmon, J., Divvala, S., Girshick, R., Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. CVPR 2016. — The comparison baseline the project deliberately departs from.

Segmentation

[9] Otsu, N. A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 9(1), 1979. — Starting point of the foreground mask; replaced for the colour path by noise-derived thresholds because Otsu placed the gap between the card body and its white label.

[10] Canny, J. A Computational Approach to Edge Detection. IEEE TPAMI 8(6), 1986. — Edge silhouette (closed colour edges, flood fill from the border).

[11] Meyer, F. Topographic distance and watershed lines. Signal Processing 38(1), 1994. — Watershed separation of touching parts after the distance transform.

[12] Zivkovic, Z. Improved adaptive Gaussian mixture model for background subtraction. ICPR 2004. — Background modelling; the empty-belt reference with a per-pixel noise map is the single-distribution, static-camera special case.

[13] Standard photographic optics on the cos⁴ illumination fall-off (e.g. Ray, S. F. Applied Photographic Optics, 3rd ed., Focal Press 2002). — Why a single global belt colour fails on a wide-angle module and the belt model must be local.

Project documents

[14] Thurmayr, A. SE-M3 — Specification v0.1 (SPEC.md), 2026. — Core goal, layer constraints, measured hardware limits, acceptance criteria A1–A15.

[15] Thurmayr, A. SE Inspect — Projektbericht Alpha 1.9.37 (PROJEKTBERICHT_ALPHA.md), 2026.