Hardware cost reduction of 30–50% is achievable when existing cameras qualify for AI inference. That qualifying condition does most of the work. It is also the part of the retrofit-first pitch that vendors who want to close quickly tend to skip.
This post covers the survey you need to run on your existing camera estate before committing to a retrofit architecture, the five technical conditions that disqualify a camera from carrying HyperQ AI Vision inference, and the economics of each scenario — cameras that qualify, cameras that need partial upgrades, and cameras that need replacement regardless of the budget position. The edge vs cloud inference post covers the network and latency side of the same deployment decision; this post covers the camera hardware side. Retrofit-first is a legitimate cost strategy. It is not always the right one.
The retrofit-first instinct and why it exists
Most factories have a CCTV estate already in place. Facilities of moderate size in manufacturing typically carry 20–60 cameras covering production floors, access points, warehouses, and outdoor perimeters. Those cameras represent a capital expenditure that has already been absorbed. The instinct to reuse them for AI inference before buying new hardware is financially rational.
The pitch from AI vision vendors often reinforces this instinct. "Works on your existing cameras" is a selling point because it reduces the headline cost of deployment. It is accurate in many cases. HyperQ AI Vision does run on standard IP cameras, ONVIF-compatible cameras, and many CCTV models that were installed for security purposes rather than inspection. The 30–50% hardware cost reduction cited in the context of retrofit deployments reflects real scenarios where the existing estate qualified and additional cameras were purchased only for specific inspection stations the existing coverage did not reach.
The problem is that "existing cameras" covers an enormous range of hardware. A 4-year-old 2MP dome camera on a 12m ceiling does not qualify for the same inference tasks as a 5MP machine-vision camera mounted 800mm from a production surface. They are both "existing cameras." Only one of them can carry AI inspection inference reliably.
The five disqualifiers: when re-use fails
Running the estate survey before scoping a retrofit deployment reveals which cameras qualify, which need partial fixes, and which need replacement. The five conditions that disqualify a camera from AI inference are frame rate, shutter type, field of view, IR sensitivity, and PoE budget. Each one has a direct effect on inspection quality that no software workaround can correct.
Frame rate insufficient for line speed
A camera running at 8fps captures one frame every 125ms. A production line moving at 0.3 seconds per unit passes a new part every 300ms. That timing looks fine — two frames per part should be enough. But the inspection system needs a clear, unblurred image at the moment the part is in the inspection window. If the part is moving and the camera's minimum exposure time is governed by its frame rate ceiling rather than its shutter speed, the result is motion blur across 40–50% of inspectable area at line speed.
The minimum usable frame rate for in-line AI inspection depends on conveyor speed and part size. For slow-moving applications (incoming inspection at a receiving dock, semi-manual assembly stations), cameras running 8–10fps are often adequate. For in-line production at 200+ parts per hour, 15fps is the working minimum, and 30fps provides reliable margin. Most CCTV-origin cameras installed before 2020 run 8–10fps at full resolution, dropping to 5fps when the network is congested.
Rolling shutter at line speed
Industrial cameras designed for machine vision use global shutter: the entire sensor captures a single instant simultaneously. Most CCTV cameras, IP cameras, and webcam-class hardware use rolling shutter: the sensor scans from top to bottom over a period of 1–10ms, meaning different rows of the image are captured at slightly different times.
On a stationary subject this is invisible. On a moving production part, rolling shutter introduces geometric distortion — diagonal lean, edge warping, horizontal banding — that varies with conveyor speed. A circular part becomes an ellipse. A straight edge becomes a curve. A defect that sits on a moving edge appears shifted from its actual position. AI inspection models trained on correctly proportioned images will misclassify geometry-distorted images at elevated rates.
Rolling shutter disqualifies a camera for inspection of parts moving faster than approximately 0.5 metres per second. For safety monitoring of workers and vehicles at normal facility speeds, rolling shutter is acceptable. For production inspection, it is not.
A process engineer on a PCB assembly line described the pattern at commissioning: the model trained cleanly on static reference images, but in production the solder joints appeared to lean slightly in one direction under rolling shutter capture. The effect was invisible at 10fps during the test; at production speed the lean exceeded the bounding box alignment the model expected. What looked like a model problem was a hardware problem. Discovering it at commissioning rather than during a customer return review is the better outcome. Discovering it during estate survey is better still.
Field of view mismatched to inspection geometry
A camera installed for security covers a wide area — typically 5–15 metres of floor at 4–6 metre ceiling height. The pixel density at that distance on a 2MP camera gives roughly 1–2mm per pixel at the inspection surface. For detecting a surface scratch 0.3mm wide on a precision component, that resolution is inadequate by an order of magnitude.
Machine-vision inspection typically requires 0.1–0.2mm per pixel resolution, which means the camera needs to be mounted 0.3–1.0 metres from the inspection surface with appropriate optics. A security camera at 6 metres ceiling height cannot be retrofitted to this geometry by changing settings. It needs a different mounting position, and in most cases a different lens or a different camera entirely.
Field of view mismatches do not generate error messages. They generate inspection models that simply cannot detect the defect classes the line requires at the defect sizes that matter. This is the disqualifier most likely to go undetected in a rushed retrofit survey — the system deploys, the model trains, and the defect that matters (a 0.2mm pin-hole in a medical device seal, a 0.3mm crack in a semiconductor leadframe) is simply too small to register reliably on the available pixel density.
Insufficient IR sensitivity for required imaging mode
Some inspection tasks require infrared illumination or near-IR imaging: detecting subsurface cracks in rubber or elastomer components, identifying delamination in composite materials, verifying adhesive coverage under opaque labels, or inspecting weld integrity in dark alloys where visible-light imaging produces insufficient contrast.
Standard CCTV cameras apply an IR-cut filter that blocks wavelengths above roughly 650nm. This is intentional — without the filter, the camera's colour reproduction under mixed illumination is poor. With the filter, the camera is effectively blind to near-IR illumination that an inspection application requires.
CCTV cameras sold for outdoor night use have switchable IR-cut filters that flip open when ambient light drops below a threshold. These work for security monitoring in darkness. They do not switch on command for specific inspection wavelengths during daylight production. Near-IR inspection applications require cameras with either a removable IR-cut filter or no filter at all — a specification that excludes most general-purpose CCTV hardware.
Inadequate PoE budget for compute at the edge
Power-over-Ethernet cameras draw between 12.95W (802.3af, Class 3) and 25.5W (802.3at, Class 4) depending on model and resolution. An existing PoE switch supporting a 20-camera estate may be operating at 60–80% of its total power budget already. Adding an edge compute device — a GPU-equipped industrial PC or edge AI appliance — to the same switch, or connecting higher-powered cameras that require 802.3bt (90W PoE) for PTZ or combined compute functions, can exhaust the switch's power allocation.
Power budget exhaustion causes cameras to cycle off and on as the switch prioritises available power, or causes cameras to underclock to stay within limits, reducing frame rates and resolution. Either outcome degrades inference quality without generating obvious error states — the system continues operating but at degraded performance that only reveals itself in elevated escape rates during audits.
Before finalising a retrofit scope, the PoE budget of every switch in the camera path needs to be confirmed against the total wattage of the existing estate plus the proposed additions. Power budget is the disqualifier most commonly missed in pre-deployment surveys because it requires IT infrastructure access that the project team often does not have at the scoping stage.
What cameras qualify for AI retrofit
A camera qualifies for AI inference retrofit when it clears all five checks: adequate frame rate for the application speed, global shutter (or rolling shutter acceptable for the application's movement speed), sufficient pixel density at the inspection distance, correct spectral response for the required imaging mode, and available PoE headroom.
In practice, the cameras most likely to qualify for AI inference without modification are:
- IP cameras installed in the last 3–4 years at 4MP or higher, positioned within 3 metres of the inspection target for a seated or slow-moving application
- Industrial cameras already installed at machine-vision distances (under 1 metre) for earlier vision programmes that were rule-based and are now candidates for model replacement
- ONVIF-compatible cameras with confirmed global shutter and 15fps+ operation at full resolution
Cameras that typically need replacement regardless of retrofit intent:
- 2MP dome cameras at 6m+ ceiling height originally installed for security surveillance
- Any camera with confirmed rolling shutter that is inspecting parts moving faster than 0.3m/s
- Pre-2019 IP cameras from unknown manufacturers where shutter type and spectral response are not documented in the datasheet
Estate survey checklist
Use this checklist before presenting a retrofit architecture to procurement.
| Camera | Location | Resolution | Frame rate (fps) | Shutter type | Mounting height (m) | IR-cut filter | PoE class | Verdict |
|---|---|---|---|---|---|---|---|---|
| [ID] | [Station] | [MP] | [fps] | Global / Rolling | [m] | Fixed / Switchable / None | 802.3af / at / bt | Qualify / Partial / Replace |
Run this for every camera in the estate before scoping begins. The retrofit percentage — how many cameras qualify outright — directly determines whether the retrofit-first architecture makes economic sense for that deployment. A facility where 80% of cameras qualify has a strong hardware-cost argument for retrofit. A facility where 20% qualify is functionally buying new hardware at full cost while paying project overhead to survey and document the estate.
Economics of each scenario
When cameras qualify: total hardware cost is the edge compute device plus any new cameras needed for coverage gaps. The 30–50% reduction relative to a full-hardware deployment is achievable and real in cases where the existing estate is 4MP or better, installed within 3 metres of inspection targets, and running global shutter sensors. This is the scenario the retrofit pitch describes.
When cameras need partial upgrades (lens swap, bracket reposition, filter removal): add 1–3 hours of on-site work per station for hardware adjustment plus camera re-commissioning. Lens upgrades for improved pixel density run $150–400 per unit. These costs are small against the full-hardware alternative but need to be scoped explicitly rather than absorbed into the contingency.
When cameras need replacement: the hardware cost approaches a full-deployment scenario. The estate survey is still worth running because it clarifies scope precisely, avoids the project risk of discovering disqualifiers during commissioning, and ensures procurement expectations are set correctly before contracts are signed. An estate survey that reveals 12 out of 20 cameras need replacement is useful information before the project starts, not during.
Frequently asked questions
Q: Can rolling-shutter cameras be used for static inspection (no line movement)? Yes. If the part is stationary during capture — on a manual inspection station, a stop-and-go conveyor, or an incoming inspection dock — rolling shutter is not a disqualifier. The distortion only appears with relative motion between the camera and the part during the exposure window. For static inspection, a rolling-shutter camera of adequate resolution and frame rate can qualify fully.
Q: What is the minimum resolution for AI surface inspection? It depends on the smallest defect that must be detected. As a working baseline: 0.1mm per pixel requires roughly 5MP at 500mm working distance with a standard 25mm lens. For applications where the critical defect is 0.5mm or larger, 2MP at that distance may suffice. The pixel density calculation needs to be run for each specific inspection geometry — there is no universal resolution floor.
Q: Do CCTV cameras installed for security ever qualify for safety AI inference? Yes, more often than for product inspection. AI safety monitoring — detecting PPE compliance, zone occupation, worker presence — operates at the scale of a human figure, not a surface defect. A 4MP camera at 6m ceiling height provides roughly 4mm per pixel at floor level, which is adequate for detecting whether a hard hat is present or absent. Many CCTV estates qualify for safety AI inference even when they do not qualify for production inspection.
Q: What is the typical on-site survey time for a camera estate assessment? A qualified engineer running a structured estate survey can assess 20–30 cameras per day, including photographing mounting positions, retrieving datasheets, and recording PoE switch configurations. A 40-camera estate typically takes 2 days on-site. HyperQ's standard deployment includes an estate survey as the first on-site activity — outputs are the completed checklist above and a revised hardware scope before any model training begins.
Q: Does the retrofit architecture affect the quality of the AI model? Camera qualification affects model quality significantly. A model trained and deployed on images with consistent lighting, resolution, and shutter behaviour produces better inference than one that must account for variable image quality introduced by borderline hardware. Deploying on unqualified cameras and then asking why the model's escape rate is elevated is a common pattern in retrofit projects that skipped the estate survey.
Your camera inventory determines whether the retrofit architecture is a genuine cost strategy or a delayed procurement exercise. Send us your camera inventory — model numbers, mounting positions, and PoE switch specs — and we will return a qualified-estate map with replacement recommendations within two weeks. No contract until the estate survey matches your inspection requirements.
Send your camera inventory for a retrofit qualification assessment