GHD Infra Developers - Recent questions tagged solutions

Deep Learning in Machine Vision Software: Benefits for Industrial Automation

Sun, 18 Oct 2026 13:05:23 +0000

A common practice is to measure the actual height variation across your part population and add 20 to 30 percent margin to account for vibration and mounting tolerance, then select aperture and working distance accordingly.

Which Cameras Actually Perform Best Across Different Industrial Tasks? There is no single "best" camera across all applications - the right choice depends heavily on speed requirements, part size, and environmental exposure. The table below compares four representative camera categories commonly specified in industrial settings, illustrating how their attributes align with different production requirements.

Standard GigE handles many high-speed applications comfortably, especially moderate-resolution inspection at cable runs beyond a few meters, but very high frame rates combined with high resolution can exceed its roughly 125 MB/s ceiling. In those cases, 10GigE, USB3, or CoaXPress interfaces provide the additional bandwidth needed, at the cost of shorter cable runs or added hardware complexity.

Code reading and traceability form a third application area that is often underestimated in scope. Pharmaceutical and food packaging lines rely on cameras to decode 2D data matrix codes at line speeds exceeding one meter per second, verifying serialization data against a database before the product proceeds to cartoning. A single unreadable code can halt a line, so camera selection for this task prioritizes depth of field and decoding software robustness over raw resolution. ClearView

Which Integration Factors Determine Real-World Reliability? Software that performs flawlessly in a vendor demo often behaves differently once connected to a plant's existing PLC network, robot controller, and historian database. Reliable integration depends on the software's native support for standard industrial communication protocols - EtherNet/IP, PROFINET, and OPC-UA chief among them - because custom-built bridges between vision software and control systems are a common source of intermittent faults that are difficult to diagnose months after commissioning. Engineers evaluating a platform should confirm not just that a protocol is "supported" on a spec sheet but that it has been deployed in a comparable line-speed environment with the exact PLC brand already running in the plant.

Industry surveys of automation deployments consistently show that inspection errors traced back to software misconfiguration or poor lens-camera matching account for a disproportionate share of unplanned downtime - some integrators estimate this figure at nearly a third of all vision-related service calls. That statistic alone explains why manufacturing engineers now treat software selection as a hardware-adjacent decision rather than an afterthought. Choosing among the available machine vision software solutions has become as consequential as selecting the sensor or lens itself, because the software layer determines whether a camera's raw resolution actually translates into usable, repeatable measurement data on the factory floor.

Lighting Design: The Component Most Often Underestimated Illumination is frequently treated as an afterthought, purchased generically rather than engineered for the specific defect type being detected, and this is a costly mistake. Structured lighting techniques - including backlighting, dark-field illumination, and diffuse dome lighting - each reveal different classes of surface and dimensional defects, and choosing incorrectly can render an otherwise excellent camera-lens combination useless for the task at hand. Backlighting, for instance, is extremely effective for measuring silhouette dimensions and detecting cracks or holes, but it provides no information about surface texture or printed markings, which require front-lit or coaxial illumination instead.

Consider a practical calculation: suppose an inspection station needs to detect a 0.2mm scratch on a metal component, and the sampling theorem requires at least two pixels across that feature for reliable detection. If the sensor has a field of view of 100mm across 4000 pixels, each pixel represents 0.025mm, giving roughly eight pixels across the scratch - comfortably above the two-pixel minimum. If the same sensor were paired with a lens that only resolves detail down to 0.05mm at the sensor plane due to poor MTF performance, the theoretical pixel count would be irrelevant because the optics themselves cannot transmit that level of detail to the sensor. ClearView

Only if the lens's resolving power, measured in line pairs per millimeter, already exceeds the requirement of the new sensor's pixel pitch. In most cases upgrading from a 5-megapixel to a 12-megapixel sensor requires a corresponding lens upgrade as well, otherwise the additional resolution simply captures blur rather than usable detail.

No, properly architected industrial deployments run inference at the edge, directly on local hardware, which avoids dependency on network connectivity for real-time decisions. An internet connection may still be used periodically for centralized model updates or performance monitoring, but production-line inspection itself should not depend on it.

How to Calculate Focal Length for Machine Vision Lenses | Technical Guide

Sun, 18 Oct 2026 12:53:35 +0000

The good news is that focal length calculation is a deterministic exercise, not a guessing game. It depends on four measurable inputs - sensor size, working distance, field of view, and required resolution - and a formula that has remained unchanged since the earliest optical systems. For teams sourcing machine vision lenses for industry, understanding this calculation removes the trial-and-error cycle of ordering lenses, testing them on the line, and returning them when they miss specification. This article walks through the formula, a worked numerical example, and the practical constraints that separate a correct calculation from one that fails once the camera is actually mounted on the machine. ClearView Machine Vision

PC-based systems, which pair one or more standard machine vision cameras with a dedicated processing unit running full vision software suites, remain the preferred architecture for complex multi-camera synchronization, deep learning-based defect classification, or applications requiring extensive image archiving for traceability. The processing ceiling on a smart camera is fixed by its embedded hardware, whereas a PC-based system can be upgraded independently of the camera hardware as algorithmic demands grow. An automotive supplier running a twelve-camera surface inspection cell, for instance, would find a PC-based architecture far more practical than twelve independent smart cameras, both for synchronized triggering and for centralized image logging tied to part serial numbers. ClearView Machine Vision

The most common causes are inconsistent lighting from ambient light changes, minor camera or lens drift, and training datasets that did not include enough variation in acceptable part appearance, causing the software to flag normal variation as a defect.

Matching Software Capabilities to the Physical Inspection Task Before selecting any platform, engineers should map the inspection task against three variables: part variability, throughput requirement, and defect scale. A pharmaceutical blister-pack line inspecting for missing tablets has low part variability and needs binary pass/fail logic running at high speed, whereas a metal casting inspection line dealing with variable surface texture benefits from software with adaptive thresholding and trainable defect libraries. Choosing an oversized software package with deep-learning classification for a simple presence check adds licensing cost and processing latency without improving accuracy.

Consider a practical sizing example. Suppose an inspection station needs to detect a 50-micron defect on a component measuring 20 millimeters across, using a sensor with a 2048-pixel horizontal resolution. Dividing the field of view by the pixel count gives roughly 9.8 microns per pixel, meaning the defect would span about five pixels - generally enough for reliable detection algorithms to distinguish it from background noise, provided contrast and focus are properly controlled. If the same sensor were used across a 60-millimeter field of view instead, each pixel would represent nearly 29 microns, and that same 50-micron defect would barely register, forcing the software into unreliable guesswork. This kind of calculation should happen before hardware is purchased, not after a system underperforms on the floor. ClearView Machine Vision

How Rugged Does the Camera Housing Need to Be? Industrial machine vision cameras are frequently marketed with IP-rated enclosures, but the rating only tells part of the environmental resilience story. IP67-rated housings protect against dust ingress and temporary submersion, which matters in washdown environments common to food processing and pharmaceutical packaging lines, but that same sealed housing can trap internal heat if the camera lacks an adequate heat-sinking design. Operating temperature range specifications deserve equal weight to ingress protection, since a camera rated for 0°C to 50°C will not survive reliably in an unconditioned warehouse inspection cell that swings toward 55°C during summer afternoons near heat-generating machinery.

Conversely, underpowered software forces engineers into constant manual retuning of lighting and thresholds whenever raw material batches change slightly, which is common in food processing and textile manufacturing. The practical test is to run the candidate software against a sample set of at least fifty known-good and fifty known-defective parts before committing to hardware, since this small pilot usually exposes whether the algorithm generalizes or simply memorizes the test set. ClearView Machine Vision

A production engineer at a mid-sized automotive parts plant once spent three weeks troubleshooting a robotic guidance cell that kept missing part edges by fractions of a millimeter. The robot was fine. The lighting was fine. The problem turned out to be a camera chosen on price alone, with a rolling shutter sensor unsuited to the line's vibration and part speed. That single procurement decision cost more in downtime and rework than the price difference between the camera purchased and the one that should have been specified from the start. Stories like this repeat across factories every year, and they are the reason a structured buying process for machine vision hardware matters more than any single spec sheet claim.

Real-Time Data Analysis via Modern Machine Vision Software

Sun, 18 Oct 2026 12:47:04 +0000

Resolving power, typically expressed through the modulation transfer function (MTF), indicates how well a lens preserves contrast at increasing spatial frequencies. A lens rated to resolve 5-megapixel sensors will not deliver sharp results on a 12-megapixel camera, even though it physically mounts and focuses light onto the sensor. Engineers should match lens resolution to sensor pixel pitch: as a practical rule, the lens must resolve at or beyond the Nyquist frequency dictated by the pixel size, or fine features will appear soft regardless of camera quality.

Coating technology deserves specific attention as well. Anti-reflective multilayer coatings reduce internal lens flare and ghosting, which matters considerably when inspection stations use strong directional lighting to highlight surface defects such as scratches or dents on reflective metal or glass components. A poorly coated lens under such lighting conditions can generate secondary reflections that obscure the very defects the system is designed to detect, effectively defeating the purpose of the inspection station. ClearView Imaging UK

Manufacturing lines that rely on human visual inspection typically catch somewhere between 60% and 85% of surface defects, depending on part complexity, lighting conditions, and inspector fatigue over a shift. Machine vision systems, by contrast, routinely achieve detection rates above 99% for well-defined defect classes once calibrated correctly, while operating at line speeds that no manual station could sustain. That gap between human capability and automated inspection is why defect detection has become one of the primary drivers behind machine vision adoption across automotive, electronics, pharmaceutical, and packaging sectors.

What Role Do Machine Vision Cameras Play in Data Timing? The camera is not a passive data source; it is an active participant in timing precision. Machine vision cameras designed for industrial use typically offer hardware-triggered exposure, GigE Vision or Camera Link interfaces with predictable bandwidth, and onboard buffering to prevent frame loss during momentary processing delays. A camera that introduces variable latency between trigger and exposure undermines the entire downstream pipeline, regardless of how efficient the analysis software is.

Rule-based algorithms remain the better choice for geometric measurements and well-defined pass/fail criteria because they are deterministic and easy to validate for regulatory documentation. Deep learning becomes worthwhile when defects are cosmetic and highly variable in appearance, but it requires a substantial labeled dataset and ongoing retraining as production conditions evolve.

Hardware acceleration plays a defining role here. GPU-based parallel processing handles convolutional operations for deep-learning inference, while FPGA-based pre-processing offloads simpler tasks like Bayer demosaicing or edge detection before data ever reaches the CPU. Some machine vision software solutions now support heterogeneous computing, dynamically assigning tasks to whichever processor-CPU, GPU, or FPGA-offers the lowest latency for that specific operation. This flexibility is particularly valuable in multi-camera cells where several inspection stations share a single processing backbone.

Custom engagements typically begin with a feasibility study using representative sample parts, including both good and defective units, run through candidate lighting and lensing configurations before any hardware is finalized. This upfront validation step is critical because switching camera resolution or lighting wavelength after a system is deployed on a live line is far more expensive than adjusting specifications during the design phase. Integrators who skip this step often discover, months into production, that their chosen resolution cannot resolve a defect class that only appears in a small percentage of parts - a costly lesson that a proper feasibility study would have caught in days rather than months. ClearView Imaging UK

How Do Global Shutter and Rolling Shutter Sensors Actually Capture an Image? A global shutter sensor exposes every pixel on the array simultaneously, then transfers the accumulated charge to a storage node before reading it out row by row. Because exposure start and stop occur at the same instant across the entire frame, a moving object is captured as a single, temporally coherent snapshot. This is analogous to a photographic flash freezing motion - every part of the subject is recorded at exactly the same microsecond, regardless of how fast it is traveling through the field of view.

What Makes a Lens Suitable for Industrial Machine Vision Applications? Selecting machine vision lenses for industry requires evaluating several interdependent parameters simultaneously rather than optimizing for a single specification. Focal length determines the field of view at a given working distance, but it must be balanced against the sensor size to avoid vignetting or underutilized image circles. A lens designed for a 1/2-inch sensor, for instance, will produce noticeable dark corners when mounted on a camera with a 1-inch sensor, because the image circle projected by the optics does not fully cover the larger imaging area.