In Part One, we discussed the first category of machine vision missed detection: the defect was not correctly imaged. That is, the camera did not see it, or the camera saw it but the image did not contain enough information. However, if the image itself is clear and the defect does appear in the frame, machine vision can still miss it. At this point, the problem enters the next stage: the camera has already “seen” it, so why does the system still fail to recognize it? The answer may come from image processing, ROI, threshold, pattern matching, algorithm, AI model, training data, trigger signal, data transmission, processing speed, PLC communication, and many other links. Therefore, in Part Two we begin our analysis after the image enters the software.

1. From Image to Decision: Pixels Are Not Understanding, and Threshold and Sensitivity Are Not Universal Keys

When a human sees a photo, they can judge based on experience: “This is a scratch.” But what a computer sees is not a “scratch,” but pixel data—large amounts of digitized pixel information. Machine vision needs to convert pixels into information that the machine can understand through image processing, feature extraction, algorithm judgment, and other steps. Therefore, a good image does not equal a good inspection result. A clear image is only one of the prerequisites for successful inspection.

Threshold is a very common method in traditional machine vision. For example, a defect can be judged when the gray value is greater than 120. This method is very effective for some inspection tasks with stable contrast, but the real production environment is not always stable. For example, the normal background originally has a gray value of 100 and the defect has a gray value of 150, making them easy to distinguish. But due to changes in lighting, material, or exposure, the background becomes 125 and the defect becomes 145. The original threshold of 120 may then begin to cause new judgment problems. Therefore, fixed rules need to be built on the basis of stable imaging.

Suppose the system misses a defect. The engineer may think of lowering the threshold, for example from 120 to 100. This may detect more defects, but at the same time, some normal textures may also be considered defects. The result is that missed detection decreases, but false detection increases. Therefore, machine vision is not “the higher the sensitivity, the better.” What is really needed is to establish a stable detection boundary under real production conditions.

2. Positioning and Region: ROI, Pattern Matching, and Product Position Changes

ROI is Region of Interest, a very common concept in machine vision systems. To reduce computation and background interference, algorithms usually do not analyze the entire image but inspect a specified region instead. This improves processing speed, inspection stability, and signal-to-noise ratio. However, if the ROI is set incorrectly, a very direct problem occurs: the defect is in the image, but not in the algorithm’s analysis area. In the end, the defect may still be judged as OK.

Suppose the system is designed so that the product center equals the ROI center, but in actual production the product shifts to the right, and the product position no longer equals the expected position. At this time, the actual defect may have moved outside the ROI. Therefore, product positioning itself is an important part of the machine vision system. Common methods include pattern matching, edge detection, feature detection, object detection, and fiducial detection. In other words, first find where the product is, then decide where to inspect.

Pattern Matching is often used to find product position or specific features. If the product rotates, translates, scales, or changes in illumination, and the algorithm’s adaptation range is limited, the matching score may be too low, ultimately causing the product not to be correctly positioned. If the subsequent ROI is based on this positioning result, the entire inspection process may be affected. Therefore, positioning errors may ultimately manifest as missed defect detection.

3. Defect Definition and Algorithm Coverage: Normal Texture, Traditional Rules, and AI Data

This is a core difficulty in many appearance inspection systems. Suppose the product surface itself has texture, grain, processing marks, color variation, and reflection, and the actual defect is close to these normal variations in color, shape, and size. Then the system needs to solve the problem of distinguishing defects from natural variation. If the rules are too strict, normal texture will be judged as NG; if the rules are too loose, real defects will be judged as OK. This is also why machine vision inspection is not simply “finding something different from the background,” but finding features that represent meaningful differences from the normal product state.

Suppose an engineer establishes a set of rules, such as scratch length greater than X, scratch width greater than Y, and area greater than Z. For typical long scratches, this set of rules may work well. But in production, small scratches, curved scratches, broken scratches, spots, cracks, and irregular defects may appear, and their visual features are different. If the rules only cover “typical scratches,” other defects may not satisfy the conditions. Thus, the defect exists, but the rule is not triggered, resulting in missed detection.

AI Vision is not “machine vision that never misses defects.” AI can also produce false negatives. Reasons may include incomplete training data, too few defect samples, new defect types not appearing in training data, changes in product appearance, changes in lighting conditions, changes in image quality, and unreasonable model thresholds. Therefore, AI changes the recognition method but does not eliminate problems with input data and system stability.

Suppose the model training data contains Scratch A, Scratch B, and Scratch C, but Scratch D appears during production. If D is significantly different from existing samples, the model may not be able to reliably judge it as a defect. This is why AI Vision projects usually attach great importance to representative training data—that is, whether the training data truly covers the variations in the real production environment. In industrial production, Good Products usually far outnumber Defective Products. For example, among 100,000 products, there may be only a few hundred real defects. As a result, AI data naturally tends to suffer from class imbalance. If the training data is not handled properly, the model may be more likely to learn “most products are normal” rather than “which subtle features indicate anomalies.” Therefore, a small number of defects does not mean the diversity of defect data can be ignored.

A deployed vision model does not necessarily maintain the same performance forever. If the production environment changes, such as product model changes, material changes, surface finish changes, lighting changes, or camera position changes, the distribution of input images may also change. Such situations can usually be understood as data distribution shift or domain shift. Therefore, AI vision systems also require continuous verification. AI models usually output a confidence score. For example, if the defect confidence is 0.62, the system needs to set what level is considered NG. If the threshold is 0.50, then 0.62 will be judged as NG; but if the threshold is adjusted to 0.70, then 0.62 will be judged as OK. Thus, a stricter threshold does not necessarily mean the system is more reliable. In practical applications, a reasonable judgment boundary must be determined based on false negatives, false positives, product risk, and inspection requirements.

4. System Real-Time Performance and Execution Chain: Speed, Trigger, Synchronization, Transmission, and PLC

Machine vision must not only be accurate, but also fast enough. For example, if the production line speed is 200 products/min, and the vision system can actually process only 100 images/min stably, then even if the algorithm itself has high accuracy, the system may still fail to inspect all products. Frame drops, missed triggers, processing queues, or processing delays may occur, ultimately causing some products to receive no complete inspection at all. The fact that an industrial camera supports 100 FPS does not mean the entire vision system can complete 100 full inspections per second. Because the system still needs to go through camera, transmission, processing, algorithm, decision, PLC, and other links. If the algorithm processing capability is only 50 images/s, then no matter how fast the camera is, it cannot truly improve the inspection capability of the entire system. Therefore, what needs attention is end-to-end throughput, not camera FPS alone.

Machine vision needs to solve a very simple but very important question: when to take the photo? This is the role of the Trigger. A typical process may be: product detected, trigger signal, camera exposure, image transmission, image processing, inspection judgment, output OK or NG. If the Trigger is too early, the product has not yet entered the optimal shooting position; if the Trigger is too late, the product has already left the optimal shooting area. In the end, the camera may indeed capture an image, but it captures the wrong image.

If the system uses multiple industrial cameras to inspect one product simultaneously, the cameras should correspond to the same product event as much as possible. If synchronization fails, for example, Camera A captures Product 001, Camera B captures Product 002, and Camera C captures Product 001 again, the system may not be able to correctly combine the inspection results. Therefore, multi-camera systems need to pay attention to trigger synchronization, exposure timing, frame ID, product tracking, and data association. Industrial cameras generate large amounts of image data. In particular, high resolution plus high frame rate brings high data throughput requirements. If interface bandwidth is insufficient, computer processing capability is insufficient, network configuration is unreasonable, or buffer management has problems, frame drops or processing delays may occur. Therefore, industrial vision systems need to consider the camera together with the interface, computer, and software as a whole.

There is another situation that is easily overlooked: the camera captures correctly, the algorithm judges NG, and the software outputs NG, but the PLC does not correctly receive the signal, or the reject mechanism does not act, and the product is ultimately released. From the production result perspective, this still appears as the defect not being effectively intercepted. Therefore, a complete visual inspection system should actually include detection, decision, communication, and action, not just image recognition.

5. Complete Troubleshooting Logic and Final Conclusion

When machine vision misses a defect, a very common solution is to switch to a higher-resolution camera. But if the problem is actually light reflection, higher pixels cannot solve it; if it is an ROI setting error, higher pixels cannot solve it; if it is insufficient AI data, higher pixels cannot directly solve it; if it is a Trigger error, higher pixels cannot solve it; if it is a PLC signal error, higher pixels are even less able to solve it. So the real question should be: Where does the inspection chain fail? Not: Which camera has more pixels?

When a system misses a defect, it can be troubleshooted layer by layer in the following order. The first layer is field of view: confirm whether the defect truly enters the camera FOV. The second layer is imaging: confirm whether the defect forms a sufficiently clear image. The third layer is pixels: confirm whether the defect has enough effective pixels. The fourth layer is lighting: confirm whether there is enough contrast between the defect and the background. The fifth layer is image processing: confirm whether ROI, threshold, and filtering are reasonable. The sixth layer is algorithm: confirm whether the rules cover this type of defect. The seventh layer is AI: confirm whether the training data covers real variations. The eighth layer is real-time performance: confirm whether camera, processing, and trigger match the production speed. The ninth layer is communication: confirm whether the inspection result is correctly transmitted. The tenth layer is execution: confirm whether the PLC, robot, or reject mechanism executes correctly. These ten layers actually form a complete machine vision inspection chain.

After Parts One and Two, we can re-understand the working process of machine vision. The first stage is to make the defect visible, obtaining a stable image through lighting, lens, and camera. The second stage is to make the image analyzable, extracting effective information through exposure, image processing, and ROI. The third stage is to enable the system to judge, recognizing defects through traditional vision or AI vision. The fourth stage is to make the judgment produce a result, executing the final action through PLC, robot, or reject system. Therefore, machine vision is not a separate “camera problem” or “AI problem”; it is a complete information chain.

When a machine vision system misses a defect, the truly effective troubleshooting approach is not to first switch to a higher-pixel camera, nor to first make the AI model more sensitive, but to start from the entire inspection chain: object, lighting, lens, camera, image, processing, algorithm, decision, action, and then confirm layer by layer. Was the defect seen? Did the image retain enough information? Did the algorithm understand correctly? Did the system process in time? Was the result correctly executed? Only by finding the link where the problem actually occurs can missed detection be solved in a targeted way. Ultimately, what a reliable machine vision inspection system pursues is not to make a single component “stronger,” but to keep the entire inspection chain stable, repeatable, measurable, and verifiable. This is also one of the most important differences between industrial vision and ordinary photography. The goal of an ordinary camera is to obtain a clear photo; the goal of an industrial vision system is to obtain data sufficient to support reliable automatic decision-making. And the true reliability of machine vision begins with this data.

Q&A for Part Two

Conclusion

Machine vision missed detection has never been a problem that can be simply attributed to one cause. It may occur outside the field of view, outside focus, outside lighting, or in algorithm rules, AI data, trigger timing, communication links, or actuators. If any link fails, a real defect may escape inspection. Therefore, the key to solving missed detection is not to rush to replace the camera with a higher pixel count, nor to blindly make the algorithm more sensitive, but to return to the complete inspection chain: Object → Lighting → Lens → Camera → Image → Processing → Algorithm → Decision → Action, and confirm layer by layer whether the defect was seen, whether the image retained information, whether the algorithm understood correctly, whether the system processed in time, and whether the result was correctly executed.

What a reliable machine vision system pursues is not the “strength” of a single component parameter, but the entire chain remaining Stable → Repeatable → Measurable → Verifiable. An ordinary camera pursues a clear photo; industrial vision pursues data sufficient to support reliable automatic decision-making. The true reliability of machine vision begins with this data.