Epson V700 Scanner: Add plate detection and colony phenotyping pipeline - #75

…r to TiffProcessor

Detect gold 3D-printed frames around agar plates using HSV
thresholding, emit plate_count/plate_boxes in metadata, and draw
green overlay rectangles on the JPEG preview. Remove integration
tests that relied on example JPG files not tracked in the repo.

Co-authored-by: Cursor <cursoragent@cursor.com>

Make plate_boxes None-able so parse_metadata() omits plate fields
when detection hasn't run (instead of silently returning zeros),
simplify overlay drawing to fill-and-restore (drops ski.draw loop),
rename converter -> processor, and mark PlateBox as private.

Co-authored-by: Cursor <cursoragent@cursor.com>

- Make detect_plates side-effect-only (results in self.plate_boxes)
- Guard against redundant detection in export_jpg
- Fix _to_rgb_uint8 redundant float round-trip for RGBA inputs
- Use round() instead of int() in _derive_dpi for robustness
- Add logging in detect_plates and warning in parse_metadata
- Document non-overlapping assumption in _draw_plate_overlays

Co-authored-by: Cursor <cursoragent@cursor.com>

New colony_detection module that operates on individual plate crops:
crop margin, optimize contrast, detect colony presence, Gaussian smooth,
and Otsu threshold. Measures per-colony area, centroid, eccentricity,
and equivalent diameter. Wired into the Lambda pipeline (auto-runs on
each detected plate, results stored in run metadata) and the CLI
(--detect-colonies flag on epson-scanner command).

Co-authored-by: Cursor <cursoragent@cursor.com>

Draw cyan contour outlines on per-plate overlay JPEGs and export a
combined colony-properties CSV (via pandas). Wired into both the CLI
(--detect-colonies now emits overlay JPEGs + CSV) and the Lambda
pipeline (uploads overlays and CSV to S3 as processed files).

Co-authored-by: Cursor <cursoragent@cursor.com>

Instead of producing separate per-plate colony overlay images, colony
contours are now drawn directly on the original export figure alongside
the plate bounding boxes before resizing.

Co-authored-by: Cursor <cursoragent@cursor.com>

Replace the 10% fractional margin with a fixed 200px crop from each
edge, which better handles varying plate sizes.

Co-authored-by: Cursor <cursoragent@cursor.com>

Have _draw_colony_contours import MARGIN_PX from colony_detection
instead of hardcoding its own default, so the crop offset stays in sync.

Co-authored-by: Cursor <cursoragent@cursor.com>

The 95th percentile was too sensitive, causing empty plates to be
misclassified as having colonies due to noise/edge artefacts.

Co-authored-by: Cursor <cursoragent@cursor.com>

…ssians

Use a DoG band-pass filter (low_sigma=0.6, high_sigma=64) followed by
10th-percentile baseline subtraction for colony detection pre-processing.
This suppresses both high-frequency noise and low-frequency background
gradients more effectively than simple Gaussian smoothing.

Co-authored-by: Cursor <cursoragent@cursor.com>

- tweak DoG parameters

Test helpers defaulted to 400x400 images which became empty after
cropping 200px from each side. Increase defaults to 600x600 and add
an empty-array guard in detect_colony_presence.

Co-authored-by: Cursor <cursoragent@cursor.com>

- minor correction to docstring

- modify display parameters

Convert all spatial colony properties from pixels to millimetres using
scan DPI (25.4 / dpi). Centroid and bbox coordinates are now
plate-relative (margin offset added back) for downstream colony picking.
Per-colony mean RGB extracted via regionprops intensity_image.

Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

…border"

This reverts commit 9464bf3.

Colonies that touch the edge of the margin-cropped region are partial
and have distorted properties. Also bump MARGIN_PX to 250 and increase
test plate sizes accordingly.

Co-authored-by: Cursor <cursoragent@cursor.com>

The JPG export draws contours from result.mask. Previously it used the
raw Otsu mask which still contained border-touching colonies that were
filtered from the properties list. Now measure_colonies returns the
post-clear_border mask so contours and properties stay consistent.

Co-authored-by: Cursor <cursoragent@cursor.com>

Replace pixel-level contour stamping with matplotlib-rendered bounding
boxes and numeric labels. Clean up inline imports (move MARGIN_PX to
top-level, use Figure/FigureCanvasAgg directly instead of pyplot with
global backend side effect) and replace BytesIO roundtrip with direct
canvas buffer read.

Co-authored-by: Cursor <cursoragent@cursor.com>

Use skimage remove_small_objects to filter noise artifacts from the
binary mask before labeling. The threshold is converted from mm² to
pixels using the scan DPI.

Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

- label after clear border

Extract run_colony_pipeline() so both the Lambda handler and CLI
delegate to one function instead of duplicating the plate-loop logic.
Fix remove_small_objects deprecation (min_size → max_size) and a
test that used a blob below the minimum-area threshold.

Co-authored-by: Cursor <cursoragent@cursor.com>