By name: featurej_documentation description: FeatureJ is a Fiji/ImageJ plugin suite for multi-scale differential feature extraction in grayscale images. This skill documents the validated Groovy automation path through the installed ImageScience classes for Derivatives, Edges, Laplacian, Hessian, Structure, and Statistics, plus the verified FeatureJ menu surface and a parameterized workflow script that saves TIFF and CSV outputs.
Install via the Fiji update site ImageScience, then restart Fiji.
Primary Use Case in This Skill Set
Open one grayscale 2D or 3D image, run one FeatureJ mode, and save the result as a deterministic TIFF or CSV file.
The most common path is Edges for a Canny-style edge map. Derivatives,
Laplacian, Hessian, Structure, and Statistics are additional validated
surfaces.
Minimal Groovy Script - Edges Mode
import ij.IJ
import imagescience.feature.Edges
import imagescience.image.Aspects
import imagescience.image.FloatImage
import imagescience.segment.Thresholder
def imp = IJ.openImage("/data/example_1.tif") // must be grayscale
def image = new FloatImage(imp)
image.aspects(new Aspects()) // pixel-unit smoothing scales
def edges = new Edges().run(image, 1.5d, true) // suppress non-maxima
new Thresholder().hysteresis(edges, 20.0d, 40.0d)
IJ.saveAsTiff(edges.imageplus(), "/tmp/example_1-edges.tif")
See GROOVY_WORKFLOW_FEATURE_EXTRACTION.groovy for a parameterized multi-mode script that saves all validated FeatureJ outputs.
Headless Fiji invocation:
/opt/Fiji.app/fiji-linux-x64 --headless --run \
/app/skills/featurej_documentation/GROOVY_WORKFLOW_FEATURE_EXTRACTION.groovy \
"inputFile='/data/example_1.tif',outputDir='/data/featurej_validation/workflow_output',mode='Edges'"
When using --run, quote File and String parameter values inside the
argument string. Unquoted paths such as inputFile=/data/example_1.tif are
parsed as SciJava expressions rather than literal file paths.
Command Quick Reference
| Mode | Validated Groovy call | Key outputs |
|---|---|---|
| Derivatives | new Differentiator().run(new FloatImage(image), 1.5d, 1, 0, 0) |
One 32-bit derivative image |
| Edges | def out = new Edges().run(new FloatImage(image), 1.5d, true); new Thresholder().hysteresis(out, 20d, 40d) |
One 32-bit image: continuous gradient magnitude if unthresholded, binary 0/255 map if thresholded |
| Laplacian | def out = new Laplacian().run(new FloatImage(image), 1.5d); new ZeroCrosser().run(out) |
One 32-bit Laplacian image, or binary 0/255 zero-crossing map if enabled |
| Hessian | new Hessian().run(new FloatImage(image), 1.5d, false) |
2 eigenimages in 2D, 3 in 3D |
| Structure | new Structure().run(new FloatImage(image), 1.5d, 3.0d) |
2 non-negative eigenimages in 2D, 3 in 3D |
| Statistics | stats.run(image, lower, upper, maskImage); stats.get(Statistics.MEAN) |
Scalar measurements for the full image or an ROI |
5 Critical Pitfalls
Pitfall 1 - RGB and indexed-color inputs are rejected
FeatureJ only accepts grayscale ImageJ image types (8-bit, 16-bit, or
32-bit). Convert RGB or indexed-color data before running:
IJ.run(imp, "8-bit", "")
Pitfall 2 - Too-small or too-large smoothing scales can make derivatives meaningless
FeatureJ's derivative-family filters are scale-space operators. Very small scales relative to the derivative order introduce sampling artifacts, while very large scales truncate the Gaussian kernel and wash out the feature.
Pitfall 3 - Edges threshold fields change the output meaning
With both threshold fields empty, Edges returns a continuous gradient
magnitude image. With one threshold, it performs hard thresholding. With two
thresholds, it performs hysteresis thresholding and returns a binary map.
Pitfall 4 - Hessian sign interpretation changes when absolute comparison is enabled
Signed Hessian eigenimages distinguish bright-on-dark from dark-on-bright structures. Absolute comparison removes that polarity information, so results from signed and absolute runs should not be mixed in the same downstream logic.
Pitfall 5 - Statistics does not support line ROIs, and float median/mode are approximate
FeatureJ Statistics accepts area ROIs but not line selections. On 32-bit
float images, the median and mode are histogram-based estimates rather than
exact values.
Parameter Tuning Quick Guide
| Observation | Fix |
|---|---|
| Derivative or edge map is dominated by pixel noise | Raise smoothingScale |
| Fine boundaries or thin ridges disappear | Lower smoothingScale |
| Edge map is almost empty | Lower the Edges thresholds |
| Edge map is too thick or full of weak responses | Raise the thresholds and keep non-maximum suppression enabled |
| Hessian or Structure output is too sensitive to tiny local texture | Raise smoothingScale and, for Structure, also raise structureIntegrationScale |
| Structure eigenimages look too spatially averaged | Lower structureIntegrationScale |
| Physical voxel calibration should control smoothing scale | Preserve calibration instead of forcing new Aspects() |
File Index
| File | Contents |
|---|---|
OVERVIEW.md |
Plugin description, image requirements, use cases, installation, limitations, and links |
GROOVY_API.md |
Validated direct Groovy API through ImageScience classes, parameter notes, ROI handling, and exclusions |
GROOVY_WORKFLOW_FEATURE_EXTRACTION.groovy |
Executable Fiji script: opens an image, runs one validated FeatureJ mode, and saves TIFF or CSV outputs |
UI_GUIDE.md |
GUI reference for each FeatureJ menu entry and the Options dialog |
UI_WORKFLOW_MULTI_SCALE_FEATURE_EXTRACTION.md |
Complete GUI walkthrough: Derivatives, Edges, Laplacian, Hessian, Structure, and Statistics on one image |
SKILL.md |
This file - LLM quick-reference card |