Imagine inspecting an assembly kit. The tray should contain five washers. This one contains four. Every washer looks perfect. No scratches, no contamination, no unusual shape. The kit is still wrong. That sounds obvious to a person. But a vision system trained to recognise unusual appearances may be answering a different question. Familiar doesn’t mean correct Consider a simple patch-based anomaly detector. It compares each image patch with a memory bank of normal examples, t

Why strided downsampling can make CNN predictions sensitive to a one-pixel shift—and what anti-aliasing can and cannot fix.

Why batching can change floating-point results—and how tiny score differences can become different inspection decisions.

Why equal Euclidean distances can hide different kinds of anomalies—and how covariance, whitening, and Mahalanobis distance change visual inspection.

Why non_blocking=True removes a CPU-side wait but does not guarantee GPU overlap—and how pinned memory, CUDA streams, and batching make the difference.

How dynamic resolution helps VLMs handle images from different cameras—and why patch counts, preprocessing, and compute budgets still matter.

High benchmark performance does not guarantee a reliable inspection system. This article presents a production-oriented framework for evaluating visual anomaly detection across thresholds, distribution shift, error costs, latency, validation, and monitoring.