Image Alignment User Guide

Applicable software: FIX Inspection Software
Document version: 2026-04-20

1. Why Image Alignment Is Needed

In real-world inspection, products from the same batch never image at exactly the same position. The reasons include:

• Mechanical positioning error: conveyors, trays, and fixtures all carry millimeter-level tolerance.
• Manual placement bias: operators place products with slight translation and rotation (typical 0° / 90° / 180° / 270° orientations).
• X-ray geometry variation: minor changes in rotary-stage angle, tilt, and magnification introduce scale differences.
• Minor posture differences within the product itself: warping and uneven thickness cause projection distortion.

The inspection ROIs, inspection rules, and reference templates drawn during Teaching are all anchored to that single teaching image's coordinate system. If the incoming image is not first "moved back" to the teaching pose, every downstream ROI will land on the wrong location, and the inspection result will inevitably be off.

One-sentence definition: Image Alignment = transform the current product image by translation, rotation, and (when necessary) scaling into the same coordinate system as the teaching template, then copy the inspection ROIs from the template into the image and run inspection on them.

2. Position in the Processing Chain

A FIX processing chain is a series of modules executed in order. A typical configuration for void inspection looks like this:

  ┌──────────────────┐
  │  Original Image  │   ← incoming X-ray image
  └────────┬─────────┘
           │
           ▼
  ┌──────────────────┐
  │ Image Alignment  │   ← this module (must come before inspection)
  └────────┬─────────┘
           │  Output: aligned image + transform matrix
           ▼
  ┌──────────────────┐
  │ Void Inspection  │   ← inspection module uses teaching ROIs
  └──────────────────┘

Key rules:

• The alignment module must sit before every inspection module that uses ROIs / teaching coordinates.
• When alignment fails, downstream inspections are flagged with an alignment error, and the current result is invalid.

3. Teaching and Runtime Workflow

3.1 Teaching phase

1. Open a typical, representative sample image (correct pose, defect-free is recommended).
2. Place the inspection ROIs and finish the inspection setup (parameters, rules, thresholds, etc.).
3. The Image Alignment module is automatically inserted before the inspection module, and the current sample image is automatically saved as the reference image. No manual configuration is required.
4. If necessary, adjust the alignment mode inside the Image Alignment module to suit the product (full-image / ROI image alignment / threshold alignment / model alignment — see Sections 4 and 5).
5. Save the teaching file — all module parameters (reference image, inspection ROIs, thresholds, model paths) are persisted together.

📌 The teaching image is critical: it should be clean, defect-free, and in the standard reference pose. Once the teaching file is released, this reference image is baked into the teaching file, and every future product will be aligned against it.

3.2 Runtime phase

1. Load the teaching file — the processing chain is rebuilt in the taught order.
2. The incoming image is fed into the chain.
3. The Image Alignment module reads the reference → computes the transform matrix → outputs the aligned image.
4. Downstream modules run inspection on the aligned image, and the ROIs naturally land in place.

4. Alignment Algorithm Overview

FIX provides 4 alignment strategies, listed from highest to lowest recommendation priority:

📌 Always try in the order ① → ② → ③ / ④. If neither ① nor ② is stable, first check the FIX project instructions to decide whether ③ applies. If that still fails, contact the FIX team to evaluate ④.

5. UI Parameter Reference

The Image Alignment panel is split into four groups from top to bottom: Selected ROI, Image Alignment Settings, Alignment Models, Threshold Alignment.

5.1 Selected ROI

When you select an ROI on the canvas, the switches below apply to that ROI.

Dynamic ROI sub-parameters

5.2 Image Alignment Settings

Controls full-image alignment (the default template-matching mode).

5.3 Alignment Models (advanced, model-based alignment)

When a product has special characteristics and the generic image-alignment functions (template matching / threshold alignment) do not perform well, switch to "advanced model alignment" — a product-specific alignment model trained and delivered by the FIX team.

Typical situations that call for model alignment:

• Large appearance variation in incoming material (color, surface finish, batch-to-batch differences cause template-matching scores to be consistently low);
• Complex background, or the product occupies only a small portion of the frame (open trays, loose-piece handling, mixed products);
• 2D → 3D deformation or occlusion (warping, tilting, partial occlusion by the fixture);
• Parameter tuning does not produce stable alignment — field tests show template-matching failure rate above 5%.

📌 If your on-site product needs model alignment, contact the FIX team: provide several representative incoming images and one teaching template image, and we will train a product-specific model and deliver a model file. Drop it into the designated folder and it will appear in the drop-down.

5.4 Threshold Alignment

This is a checkable group — checking the group title enables threshold alignment mode, which overrides the default template matching.

Principle: apply the same binarization to reference and current images → filter contours → compute the foreground centroid → translate the current image so its centroid matches the reference centroid. Only translation is handled (plus an optional 180° flip); rotation and scale are not.

When to use:

• The product only has translational offset (fixture is accurate, only a few millimeters of drift);
• Product-vs-background grayscale contrast is very clear (gaskets, small chips, stand-alone parts);
• Extreme speed is required — threshold alignment is 3–5× faster than template matching.

6. Alignment Flow in Detail — How the Matching Score Is Computed

This section explains the full computation for the default mode (template matching). Understanding it helps in setting the "Alignment Score <" threshold sensibly.

6.1 Overall steps

 ┌─────────────────────┐   ┌─────────────────────┐
 │   Reference image   │   │   Current image     │
 │   (saved at teach)  │   │   (incoming)        │
 └──────────┬──────────┘   └──────────┬──────────┘
            │                         │
            └───────────┬─────────────┘
                        ▼
              ┌──────────────────────┐
              │ Step ① Size match    │
              │ Sizes differ?        │
              │ Pad / crop current   │
              │ to reference size    │
              └──────────┬───────────┘
                         ▼
              ┌──────────────────────────┐
              │ Step ② Object finding    │ ← can be skipped via
              │ Use contour/threshold    │   "Disable Object Finding
              │ to find product bbox     │    Before Alignment"
              └──────────┬───────────────┘
                         ▼
              ┌──────────────────────────┐
              │ Step ③ Rotation detect   │ ← can be skipped via
              │ 90° / 180° / 270° ?      │   "Disable Auto Rotation"
              │ If yes, rotate back      │
              └──────────┬───────────────┘
                         ▼
              ┌──────────────────────┐
              │ Step ④ Template      │ ★ score is produced here
              │ matching:            │
              │ slide template over  │
              │ current, compute     │
              │ per-pixel similarity │
              └──────────┬───────────┘
                         ▼
              ┌──────────────────────┐
              │ Step ⑤ Transform     │
              │ From the best match  │
              │ position, derive the │
              │ 2×3 translation /    │
              │ scale / rotation mat │
              └──────────┬───────────┘
                         ▼
              ┌──────────────────────┐
              │ Step ⑥ Resample      │
              │ Warp current with    │
              │ the transform →      │
              │ aligned output image │
              └──────────┬───────────┘
                         ▼
              ┌──────────────────────┐
              │ Step ⑦ Score evals   │
              │ • Alignment Score    │
              │ • SSIM Score         │
              │ Below threshold →    │
              │ failure              │
              └──────────────────────┘

6.2 How the template-matching score is computed (Step ④ detail)

Template matching uses normalized cross-correlation. The reference image becomes a "template" that is slid over the current image, position by position, and compared:

Current image (search area)
 ┌──────────────────────────────┐
 │                              │
 │    ┌──────┐                  │
 │    │tmpl  │←── compare pixel │
 │    │      │    similarity    │
 │    └──────┘    at this spot  │
 │         ┌──────┐             │
 │         │tmpl  │←── next     │
 │         └──────┘    position │
 │              ...             │
 └──────────────────────────────┘
        ↓
  Every position yields a score
        ↓
  Position of the highest score
         = best match

The score at each position = the normalized correlation coefficient between the template pixels and the search-image pixels at that position, in the range −1 ~ +1 (mapped inside the software to an integer in the range 0–99999 depending on implementation):

Alignment Score = the highest score found across the image.

• Default threshold 1500 (corresponds to a fairly good correlation).
• Only when "highest score ≥ threshold" is alignment considered successful.

SSIM Score is computed after alignment, between the aligned image and the reference image (0–1):

• ≥ 0.85: excellent alignment quality;
• 0.60 ~ 0.85: acceptable;
• < 0.60: structures are clearly off — adjust parameters.

6.3 Rotation detection flow

 Reference + Current
        │
        ▼
 ┌────────────────────────┐
 │ "Use Enhanced Rotation │
 │  Detection" checked?   │
 └────┬───────────┬───────┘
      │ Yes       │ No
      ▼           ▼
 ┌─────────┐   ┌───────────────┐
 │Keypoint │   │ Rotate current│
 │extract  │   │ by 0° / 90° / │
 │& match  │   │ 180° / 270°   │
 │→ angle  │   │ and run       │
 └────┬────┘   │ template match│
      │        │ on each;      │
      │        │ pick best     │
      │        └──────┬────────┘
      └────────┬──────┘
               ▼
   Determined product rotation
               ▼
     Rotate current back to 0°
               ▼
       Proceed to Step ④

6.4 Dynamic ROI secondary fine-alignment

After global alignment lands the whole product, individual small components (QFNs, single capacitors, locating pins) may still carry residual offset. For ROIs ticked as "Set as Dynamic ROI", the algorithm performs a second local tracking on top of the global alignment:

  Global alignment done (body landed)
                │
                ▼
  ┌────────────────────────────────┐
  │ For each Dynamic ROI:           │
  │                                 │
  │  ① Center on the taught ROI     │
  │  ② Expand outward by            │
  │     "Search Padding" pixels     │
  │  ③ Run template match again     │
  │     inside the expanded area    │
  │  ④ Best score ≥ threshold?      │
  │      Yes → move the ROI to the  │
  │            best match location  │
  │      No  → keep original spot,  │
  │            write score to UI    │
  │  ⑤ If "Move Child ROI" is       │
  │     checked, children follow    │
  │     the parent as a group       │
  └────────────────────────────────┘

7. Choosing an Alignment Method

8. Troubleshooting

Debugging tips

1. Check "Show Alignment Score" to observe the actual score and judge if the threshold is reasonable.
2. Watch the SSIM Score: ≥ 0.85 = excellent alignment; 0.60–0.85 = acceptable; < 0.60 = reconsider parameters.
3. The module preview window shows "current | reference | aligned" side by side — visual inspection already tells you the alignment quality.
4. When using model alignment, press the "Test" button first to confirm the detection box is correct, then release to production.

9. Performance and Deployment Notes

• Threshold Alignment is the fastest (millisecond-level), template matching is next, model alignment is the slowest but the most robust.
• The GPU edition of FIX automatically accelerates model alignment and parts of template matching.
• The reference image path stored in the teaching file must be kept in sync when deploying to another machine. Recommend placing the reference image and the teaching file in the same relative directory, or on a shared network path.

10. Related Documents

• Image Alignment entry on the official help site: https://help.floatech.com/docs/HbBOXa
• Case collection: https://help.floatech.com/docs/cases