No. Standard CMOS and CCD sensors used in visible-light cameras are built on silicon photodiodes that have very low quantum efficiency above roughly 1000 nm, so adding an SWIR bandpass filter to a silicon-based sensor mainly blocks light without producing a usable image. A dedicated InGaAs sensor is required to achieve meaningful sensitivity across the 900-1700 nm range used in wafer transmission imaging.
Quality control in electronics assembly represents another area of strong adoption, since defects like tombstoned components, insufficient solder paste, or misaligned connectors are visually inconsistent and difficult to define through fixed rules. Automotive weld inspection, packaging verification, and textile flaw detection follow a similar pattern: whenever defects vary in appearance, scale, or position, a learned model tends to outperform a rules-based one. It’s worth noting, however, that not every application needs this level of sophistication – a system counting discrete objects on a conveyor or verifying the presence of a barcode is often better served by simpler, faster classical algorithms that consume less computational overhead and are easier to validate for regulatory documentation. machine vision Components
How Does Deep Learning Actually Change Image Analysis on the Factory Floor? Traditional machine vision systems inspect images using algorithms like edge detection, blob analysis, and pattern matching, all of which require precise calibration for each new part or defect type. Deep learning models, particularly convolutional neural networks, instead learn hierarchical features directly from training images: edges and textures in early layers, shapes and part-specific structures in deeper layers. This layered feature extraction allows the software to recognize subtle anomalies, such as hairline cracks in cast metal components or inconsistent solder joints on a printed circuit board, without an engineer manually specifying what those defects look like in pixel terms.
EtherNet/IP and the Implicit vs Explicit Messaging Choice EtherNet/IP remains the default choice in North American plants running Rockwell or Omron controllers. Explicit messaging is simple to configure and works well for non-time-critical data such as inspection statistics or recipe downloads, but it introduces variable latency because it rides on standard TCP/IP request-response cycles. Implicit messaging, by contrast, uses pre-configured connections with a fixed Requested Packet Interval, delivering data with much tighter, more predictable timing-often under 5 milliseconds-which matters enormously for robotic guidance applications where a stale coordinate can mean a missed pick.
Ongoing costs differ in nature rather than necessarily being higher overall: classical systems incur recalibration costs when parts or conditions change, while deep learning systems incur periodic retraining and drift-monitoring costs. For high-mix production environments, deep learning maintenance is often less burdensome, while for stable, single-part lines, classical system maintenance can be simpler and cheaper.
A straightforward single-camera barcode verification station can often be commissioned in a few days, while a custom multi-camera cell with machine learning classification may take several weeks to months including data collection and validation. Timelines depend heavily on how much sample data and defect documentation is available upfront.
Many integrators mistakenly believe a faster frame rate alone solves this problem. Frame rate determines how many images are captured per second, but it says nothing about how long each individual exposure lasts. A camera running at three hundred frames per second with a five-millisecond exposure will still blur fast-moving targets badly, because the exposure time, not the frame rate, is what freezes motion. This distinction is the single most misunderstood concept among teams troubleshooting blur on existing industrial machine vision cameras deployments.
The practical trade-off is computational: a learned depth model typically requires a GPU or dedicated inference accelerator to hit the sub-100-millisecond latency that a robotic pick cycle demands, whereas classical stereo can often run on a CPU or FPGA within similar time budgets. Many integrators now deploy a hybrid approach, using geometric triangulation as the default and falling back to a learned model only for regions flagged as low-confidence, which keeps overall latency predictable while still handling difficult surface finishes. machine vision Components
In many cases, yes, provided the existing camera resolution and lighting consistency meet the minimum requirements for the defects being targeted. However, if the current hardware was selected for a rule-based system with different image quality demands, an upgrade to higher-resolution sensors or more uniform lighting may be necessary to achieve acceptable model accuracy.
The fix is rarely exotic. Engineers typically implement a strobe-and-acknowledge sequence: the PLC raises a trigger bit, the vision system latches the image, processes it, and holds its result bits stable until the PLC explicitly acknowledges receipt with a handshake bit of its own. This pattern, sometimes called a four-wire handshake even when implemented entirely over a fieldbus rather than discrete I/O, eliminates race conditions almost completely. Custom machine vision systems built for high-speed sorting lines almost always use this pattern rather than relying on simple level-triggered logic.