The embedded vision preprocessing co-design framework (EVP-CDF): Methodology and decision procedure for real-time detection pipelines

Abstract

Preprocessing in embedded object detection pipelines is often treated as a fixed implementation detail, despite its dual role in shaping the detector input distribution and contributing directly to end-to-end latency, throughput, and timing stability. This paper presents the Embedded Vision Preprocessing Co-Design Framework (EVP-CDF), a reusable methods framework that formalizes preprocessing as a constraint-aware co-design space for embedded one-stage detectors. EVP-CDF integrates three coupled levers: training-time photometric augmentation selection, boundary input scaling strategy selection (normalization versus standardization variants), and inference pipeline optimization with explicit enforcement of training–inference preprocessing alignment. Two method artifacts operationalize the framework: (i) the Preprocessing Alignment Matrix (PAM), which provides a taxonomy of mismatch types (statistical, geometric, photometric, and implementation-level) and their expected deployment failure signatures, and (ii) an evidence-guided decision procedure that selects a feasible pipeline under accuracy, latency, throughput, memory, and jitter constraints while returning an auditable alignment checklist. A minimal reporting specification (MR-Spec) is additionally proposed to standardize preprocessing disclosure, timing breakdowns, tail latency and jitter reporting, and deployment context documentation. All quantitative evidence used to support the framework is consolidated primarily from three prior peer-reviewed studies and is complemented by a targeted cross-domain ablation on COCO128 that isolates preprocessing mismatch effects without retraining or dataset modification.