StomataHub Detector

A planned web-based detector + review workspace for end-to-end stomatal analysis: upload microscope images, run models, verify/correct predictions, and export measurements with full provenance.

Phase 1 details Phase 2 details Share feedback

Upload

Single images or batches; optional calibration (µm/px) and metadata tags.

Detect

Run a selected model; tile large images; apply NMS/thresholding for stable outputs.

Measure + export

Counts, density, size distributions, per-instance tables; export with provenance.

Phase 1 Detection Batch processing Metrics + exports

Phase 1: Batch detection + measurements

Phase 1 is designed for practical throughput: upload batches, run inference, and get a clean export that can go straight into analysis scripts or spreadsheets. The goal is not just a picture with boxes, it's a reproducible measurement output, with model/version/settings attached.

Input handling

Formats: JPG/PNG/TIFF (multi-page TIFF is a target).
Calibration: µm/px for pixel→physical conversion; optional field-of-view to compute density.
Pre-processing: configurable normalization (e.g., CLAHE) with settings recorded in the run report.
Large images: overlapping tiling with stitched outputs to avoid edge artifacts.

Inference + post-processing

Model families: object detectors (e.g., YOLO-style) for bounding boxes; mask-based models added when available.
Thresholding: user-selectable confidence thresholds with transparent defaults.
NMS: non-maximum suppression to resolve overlapping predictions consistently.
Filters: optional size/shape filters (useful for removing debris-like false positives).

Measurements (planned outputs)

Per-image: total count, confidence distribution, size distribution summary.
Per-instance: box geometry, area estimate, and (where available) mask-derived area/shape traits.
Density: stomata/mm² when scale + field-of-view are known.

Export formats

Export	What you get
CSV	One row per detected stoma + measurement columns (easy for R/Python/Excel).
JSON	Structured run report (model + settings + per-image summaries + per-instance records).
Annotation export	COCO-style JSON is a target; other formats may be added based on user demand.

Example per-instance record (illustrative)

{ "image": "sample_001.tif", "stoma_id": 17, "bbox_xyxy_px": [412, 238, 468, 298], "confidence": 0.93, "area_px2": 3240, "area_um2": 12.96, "notes": {"model": "shd-detector-v1", "nms_iou": 0.5, "conf_thr": 0.25} }

Phase 2 Deblurring Denoising Illumination correction

Phase 2: Image restoration (deblurring/denoising)

Phase 2 is a restoration pipeline for difficult images, motion blur, defocus blur, sensor noise, compression artifacts, and uneven illumination. This is not a “make it prettier” filter: the goal is to improve signal-to-noise while preserving biologically meaningful structure so downstream detection and measurement become more reliable.

Restoration tasks (planned)

Deblur: learned deconvolution-style models (U-Net style) with optional blind blur estimation.
Denoise: noise-aware restoration tuned for microscope-like noise (Poisson + Gaussian).
Normalize: illumination correction / background flattening to reduce vignetting-like effects.

Training/evaluation notes (planned)

Paired data: where possible, train with clean/blurred pairs (including synthetic degradations).
Metrics: PSNR/SSIM for paired benchmarks; task-based metrics (detection F1 / count error) for real-world validation.
Artifact awareness: flag over-smoothing or hallucination risk; keep original available for side-by-side comparison.

Outputs + downstream comparison

Restored images: downloadable outputs plus a parameter record (method + strength + model version).
Difference maps (target): optional “what changed” overlays for QC.
Optional re-run: run Phase 1 detection on both original and restored images to compare counts/measurements.
Batch summaries: per-batch before/after charts (target) for count error and confidence shifts.

Interactive review + export

The detector is intended to be interactive: you should be able to validate outputs, correct mistakes, and export an updated dataset without the page fighting your workflow.

Review: browse detections across a batch; filter by confidence; spot-check outliers.
Correct: adjust boxes/masks, remove false positives, add missed instances.
Recompute: update counts/density/size distributions after edits.
Export: download updated annotations + measurements + provenance (model/settings).