Haystack Imaging
The Situation
The goal was to measure surface quality traits in apple varieties - russeting patterns, colour uniformity, early bruising - using a push-broom hyperspectral camera. A push-broom sensor captures one line at a time as the object moves beneath it. For flat samples this is straightforward. For a sphere, it is not: a single pass captures at most half the surface, the edges appear darker than the centre due to Lambert's cosine law, and any wobble in the rotation axis shows up as misaligned stripes in the final image.
What was needed was a device that would rotate apples at consistent, precise angles; synchronise with the hyperspectral sensor's acquisition trigger; and produce images clean enough to stitch into a seamless, calibrated map of the whole fruit. Off-the-shelf rotation stages existed, but none could interface with the sensor trigger, handle the irregular geometry of real fruit, or be quickly adapted as requirements changed.
The Approach
We designed and built the Apple Rotator Device (ARD) from scratch. The mechanical assembly was designed in CAD and 3D-printed in black PLA - black to absorb stray reflections that would contaminate the spectral signal. Two stepper motors drive a clamping mechanism that holds the apple at both poles, allowing full 360° rotation with 0.176° increments. Four small white tracking dots bonded to the fixture act as fiducials: the post-processing pipeline detects them in every frame and uses their positions to correct any rotational drift or mechanical jitter automatically.
An Arduino Nano manages the motor sequence and monitors a light sensor that detects the hyperspectral camera's illumination pulse - so the rotation advances precisely in step with each acquired line, regardless of the sensor's internal timing. The device mounts on a standard ACRA-SWISS rail, making it compatible with existing camera stands without modification.
The post-processing pipeline runs in Fiji with Python automation. It ingests the frame sequence, detects and corrects fiducial drift, extracts the vertical centre slice from each frame (the only column free from Lambert's cosine law shading), and stitches those slices into a continuous 2D unrolling of the whole surface. A reference image of a uniform white cylinder, captured before each session, normalises brightness across the scan, removing residual shading without distorting spectral values.
The Result
The device delivers a complete, calibrated spectral map of every apple surface, all 360°, in a single automated scan. The fiducial correction eliminates the jitter visible in uncorrected scans, and the shading normalisation removes the edge darkening that made previous hemisphere-only images unusable for quantitative analysis. The output is a standard hyperspectral data cube with known spatial coordinates tied to the fruit surface, ready for downstream machine learning.
The ARD also works with a DSLR for RGB scouting scans, making it a general-purpose whole-fruit imaging station. It is now used routinely as part of a variety evaluation workflow.
