Edges Before Embeddings: A Confidence-Aware Blur Gate for Vision-Language Pipelines

Researchers present MagikaDocumentFromPixel, a lightweight CPU-based image quality gate that detects blur in vision pipeline inputs within 7ms, preventing wasted compute on downstream tasks. The system achieves 98.03% F1 score using MobileNetV3-Large with an Edge Prior Module, establishing a reusable design pattern for production vision systems.

This research addresses a practical problem in production vision systems: blurry inputs silently degrade performance across OCR, retrieval, and vision-language models, wasting computational resources on unrecoverable outputs. The proposed blur detection gate acts as a gatekeeper, filtering problematic inputs before they reach expensive downstream processes. The technical contribution centers on identifying input resolution as the dominant performance lever—showing that architectural capacity only becomes beneficial at 384px or higher—and introducing the Edge Prior Module, which provides the network direct spectral evidence of blur patterns that classical heuristics rely upon, yielding a +1.3 F1 improvement.

The work demonstrates the value of systematic empirical search across 46 configurations with 8 sweeps, moving beyond traditional architecture-focused optimization to identify domain-specific bottlenecks. By grounding the solution in classical selective prediction theory, the authors create a confidence-aware routing framework that extends beyond blur detection. The observation that this design pattern recurs across Magika content-type detection, risk-controlled VLM pipelines, and DocVLM suggests a generalizable approach to quality gating in vision systems.

For practitioners building production vision pipelines, this represents a concrete efficiency gain: a 17MB ONNX model running on CPU achieves near-perfect classification performance without requiring GPU acceleration. However, the authors transparently acknowledge limitations—single motion-blur distribution, single-seed results, and qualitative rather than measured calibration—which constrains generalization claims. The work exemplifies responsible ML research by explicitly defining scope boundaries rather than overstating applicability.

• →Input resolution dominates blur detection performance, with architectural improvements only materializing at 384px or higher
• →Edge Prior Module adds spectral evidence channel to networks, improving F1 by 1.3 points and establishing a reusable pattern
• →7ms CPU-based blur gate prevents costly downstream OCR and VLM calls on unrecoverable inputs
• →Systematic 46-configuration empirical search identified domain-specific optimization opportunities beyond standard architecture tuning
• →Authors openly acknowledge limitations including single motion-blur evaluation and single-seed results, avoiding overstatement