microdefectcv

Method Overview — MicroDefectCV

This document describes the internal pipeline of the detect_defects function.

---

1. Grayscale Conversion

If the input image is colour (BGR), it is first converted to grayscale using cv2.cvtColor. All subsequent processing operates on a single-channel uint8 image.

2. SEM Metadata Bar Removal

SEM images frequently contain a bottom annotation bar (scale bar, metadata text). The pipeline scans the bottom 18 % of the image for rows whose median intensity is below 20 or whose column standard deviation is below 5. The first such row defines the crop point.

3. Mode / Profile Selection

The algorithm supports six named profiles:

Mode	Target defects	Grain boundary suppression	Needle detection
pbi2	Bright PbI₂ particles	No	Yes
pinhole	Dark pinholes	No	No
2d	Both	No	No
3d	Both	Yes	Yes
3d_2d	Both	Yes	Yes
auto	Determined from image stats	Conditional	Conditional

In auto mode the following image statistics are computed to select a profile:

• Canny edge density
• Mean Sobel gradient magnitude
• Top-hat bright spot density

4. Preprocessing

1. Gaussian blur — reduces high-frequency noise before CLAHE.
2. CLAHE (Contrast Limited Adaptive Histogram Equalization) — boosts local contrast so that subtle defect boundaries become visible. Clip limit and tile size are tuned per mode.

5. Grain Boundary Suppression (3D / 3D-2D modes)

In 3D perovskite images, grain boundaries cast pseudo-defect shadows. A suppression mask is constructed by:

1. Canny edge detection on the enhanced image.
2. Dilation of edges to form thick boundary zones.
3. High-gradient pixel thresholding (top 20 % of Sobel magnitude).
4. Union of both masks with morphological closing.

Any candidate defect overlapping this mask is discarded.

6. Bright Particle Detection (PbI₂)

1. Multi-scale Top-Hat transform using elliptical kernels (sizes vary per mode). This isolates bright regions smaller than the kernel size.
2. Dual percentile thresholding (high and low) on the combined Top-Hat response.
3. A fill-ratio guard prevents false over-segmentation.
4. Morphological open + close to clean the mask.

7. Dark Pit Detection (Pinholes)

1. Percentile thresholding on the enhanced image: pixels below the N-th percentile (dark_percentile) are marked.
2. A fill-ratio guard triggers a stricter percentile if too much of the image is flagged.
3. A separate micro-threshold at the 3rd percentile captures very small pinholes.
4. Morphological close + open to smooth boundaries.

8. Needle Crystal Detection (PbI₂ Needles)

1. Large rectangular Top-Hat transform (11×11 rect kernel) isolates elongated bright structures.
2. Elongated morphological closing (3×1 kernel) bridges gap artefacts in long needles.
3. Contours are filtered by minimum aspect ratio (length/width ≥ 2.5 by default).

9. Contour Filtering

For each candidate contour, shape features are computed:

• Area — rejects tiny noise and oversized regions.
• Circularity — 4πA/P² (dynamically relaxed for large dark pits).
• Solidity — area / convex_hull_area (rejects irregular shapes).
• Aspect ratio — rejects very elongated blobs in bright/dark paths.
• Interior vs exterior mean intensity — ensures correct polarity (bright/dark).
• Contrast — |ext_mean − int_mean| / ext_mean.

A confidence score is computed from these features and filtered against a threshold.

10. Non-Maximum Suppression (NMS)

Overlapping detections are deduplicated using IoU-based NMS (intersection-over-union threshold 0.4 by default), retaining the highest-confidence detection.

11. Output

The function returns a dictionary containing:

• mask — binary uint8 mask
• enhanced — CLAHE image
• defect_count — integer
• defect_area_ratio — float in [0, 1]
• contours — list of OpenCV contours
• mode — resolved mode string
• intermediates — per-stage images (when return_intermediate=True)

This site is open source. Improve this page.