The same product, which tested normal yesterday, suddenly started missing detections today.

The threshold hasn't been changed, the program hasn't been modified, and the product looks pretty much the same—but as a result, it’s unstable.

The most common on-site approach is to keep adjusting the threshold, adding filters, tweaking the exposure, or even retraining the model. After all this back-and-forth, the current batch of samples finally looks fine—but as soon as you switch to a new batch of products, the problem reappears.

Many times, the problem isn't with the algorithm at all.

Rather, the target was never consistently captured from the very beginning.

Whether or not detection seems to simply determine “yes” or “no,” what truly determines stability is whether these two states can produce a clear and stable distinction in the image.

Today, the target is black; tomorrow, it turns gray. A slight change in angle and the outline disappears. As soon as the surface’s reflection changes, the original threshold immediately becomes ineffective.

With this kind of image, even the most sophisticated algorithms can only offer a barely adequate remedy.

The component outlines are unstable—stop obsessing over the threshold.

When inspecting components on flexible circuit boards, the biggest concern is that the target and background might get mixed up.

If the detection is truly focused on the component outlines, backlighting is often a more direct approach than repeatedly adjusting the exposure.

Backlight doesn't care about the color of the surface, nor does it need to analyze complex textures—it simply turns the target into a clear black-and-white silhouette.

Once the contour has stabilized, simpler methods such as area, width, and connected components become easier to implement for stable detection.

Severe reflection doesn't necessarily mean the light is too bright.

FPC flexible boards, solder joints, and metal components are prone to producing highlights.

If the sample’s position shifts even slightly, the reflective area will move accordingly, and edges that were once sharp may suddenly become blurred.

Many people immediately reduce the exposure as soon as they see reflections.

As a result, although the highlights became weaker, the targets also dimmed together, and in the end, no one could see clearly anymore.

What really needs to be adjusted is the direction in which light enters the lens.

Strip-side lighting is ideal for highlighting edges, height differences, and slender structures. By adjusting the illumination angle, you can make the target features stand out while minimizing direct reflections entering the lens.

So, the reflection issue isn't necessarily due to light being too intense—it could also be because the light is shining in the wrong direction.

A bright image doesn't necessarily mean that blemishes are very noticeable.

With flexible, semi-transparent materials like masks, a common issue is overexposure in bright areas while dark areas lack detail.

When detecting dirt, the key isn't to make the mask appear whiter; rather, it's to create a stable contrast between the stains and the normal areas.

If the stain affects the light transmittance, backlighting can amplify this difference.

However, backlighting isn't suitable for all types of stains. If the defect primarily manifests as changes in surface color or texture, consider using front diffuse light, coaxial light, or low-angle light.

Before selecting a light source, you need to first clarify:

Does the defect alter the contour, reflection, transmission, or surface texture?

For complex products, don't leave all the details to the algorithm.

The clutch disc surface has a complex structure and is further complicated by metallic reflections, textures, and oil stains.

But what the project really needs to determine may just be whether the rivets and springs have been installed.

Since the detection focuses on the structural outline, there’s no need for the algorithm to analyze every surface detail.

After using the backlight, colors, oil stains, and highlights are muted, while rivets, hole positions, and springs produce more stable outlines.

A good lighting setup isn't about making the picture look prettier.

Instead, it transforms complex problems into ones that algorithms can easily evaluate.

To determine on-site whether a solution is stable, just look at three things.

First, have the goal and the background truly been separated?

Second, after changing the location, batch, and time, can the target feature still appear repeatedly?

Third, does the light source enhance the target, or does it also amplify irrelevant textures and reflections?

A picture looks good but doesn't mean anything.

Only when the results remain stable after taking dozens of consecutive shots and testing multiple batches of samples can we say that the solution is truly viable.

If the detection results are unstable, don't rush to switch models, add filters, or adjust thresholds.

First, ask yourself:

Is the same goal really depicted identically in every single image?

If the answer is no, the problem most likely lies not with the algorithm but with the first image.