Sigit Dwi Nugroho

Why I built this

Avocado quality is usually checked the old way: squeeze it, cut it open, or punch it with a penetrometer. All of those are subjective or destructive, and none of them scale for a buyer or a processor who wants to know now whether a fruit is ready. For my undergraduate thesis at Padjadjaran University I asked a simple question: can a small, portable device tell whether an avocado is ripe without touching or cutting it?

My answer was a prototype based on Laser-Light Backscattering Imaging (LLBI). A low-power laser shines onto the fruit, light scatters inside the flesh, and a camera watches the glowing spot. How far and how evenly that light spreads changes as the fruit ripens, so the scattering image itself becomes a ripeness signal.

Finding the method

Before designing anything I spent weeks on literature: a 2024 review of LLBI on agricultural products, studies of NIR penetration depth in Hass avocados, Haralick's 1973 texture features, and the MobileNetV2 paper. Two decisions came out of that reading and shaped the whole project:

1. Use two wavelengths. Red (650 nm) scatters near the surface, while near-infrared (850 nm) penetrates deeper through the thick exocarp, so I built the prototype to work with both and let the data decide which one matters.

2. Compare two very different classifiers. A classical pipeline with physics-informed features (gradient slope + GLCM texture, classified by an SVM) against a deep-learning approach (a MobileNetV2-based CNN), instead of assuming deep learning would automatically win on a small dataset.

Designing the prototype: from AutoCAD to Inventor

The mechanical design started as 2D drawings in AutoCAD, then moved to Autodesk Inventor where the whole assembly came together. The body is built around two dark chambers that isolate the camera from ambient light: an upper chamber with the LCD, control buttons and the laser window, and a lower chamber that holds the fruit on a custom holder aligned with the laser and the camera module.

Labeled exploded view of the prototype: LCD, buttons, upper and lower chambers, Raspberry Pi 4B, camera module, 850 nm laser module, and avocado holder — Exploded view: a. LCD, b. buttons, c. upper chamber, d. Raspberry Pi 4B, e. lower chamber, f. Camera Module V3, g. 850 nm laser module, h. avocado holder. The shells and holders were 3D-printed from STEP exports.

Everything was modeled iteratively (the Inventor version history tells the story of many shell revisions), vendor models for the Raspberry Pi 4B, the Sony IMX708 camera module and the 850 nm laser module were imported to check fits, and the final parts were exported as STEP files for 3D printing.

Electronics: from a sketch to a PCB

I sketched the wiring first in Fritzing, then redesigned it properly in KiCad as a schematic and a small 2-layer interface board. The board carries three push buttons for the kiosk controls, a transistor-driven circuit (MOSFET IRLZ44N) that switches the 850 nm laser with hardware PWM, an LCD connector, and mounting holes that line up with the Raspberry Pi as a Pi-HAT. Designing my own PCB, rather than wiring everything loose, was one of the parts of the thesis I am proudest of.

Fritzing breadboard diagram: laser module, MOSFET, buttons, Raspberry Pi 4B, camera module, LCD, and fan — The full wiring diagram: 850 nm laser module driven by a MOSFET, three push buttons, the Camera Module V3 over CSI, the 3.5" LCD over HDMI, and a cooling fan.

KiCad PCB layout of the Pi-HAT interface board for avocado ripeness classification — The KiCad PCB layout, a Pi-HAT styled interface board: LCD, laser driver, and the three buttons.

Building the dataset: 110 fruits, 14 days, two lasers

Five avocado ripeness stages from firm to ripe — The ripeness stages used in the study, tied to penetrometer firmness. The binary classification task splits at 67 N: Mentah (unripe, ≥ 67 N) vs Matang (ripe, < 67 N).

Raw 650 nm backscattering image on an avocado — Raw capture, 650 nm (surface-focused).

Raw 850 nm backscattering image on an avocado — Raw capture, 850 nm (deeper NIR scattering).

Backscattering comparison between unripe and ripe avocados at 650 nm and 850 nm — Mentah (unripe) vs Matang (ripe) at both wavelengths: the scattering footprint changes with ripeness and gets sharper at 850 nm.

Machine learning: SVM vs CNN

Both classifiers ate the same preprocessing pipeline (grayscale, Gaussian smoothing, thresholding, centroid-crop to 400×400, CLAHE, sharpening, dilation):

The SVM pipeline compressed each image into six physics-informed features: the gradient slope between two annular rings of the backscattering blob, plus five GLCM texture features. The CNN pipeline fine-tuned a MobileNetV2 backbone (ImageNet weights, frozen) with a small classification head on 150×150 inputs. Both were evaluated with a leakage-free nested 5-fold protocol and then judged once on the 16-fruit external holdout.

Preprocessing pipeline stages from original capture to final preprocessed image — The preprocessing chain every image went through before feature extraction.

Model	CV (94 fruits)	External holdout (16 fruits)
SVM, 850 nm (slope + GLCM)	84.04% (AUC 0.914)	87.5% (AUC 0.900)
CNN, 850 nm (MobileNetV2)	67.02% (AUC 0.750)	56.25%
SVM, 650 nm	51.06%	37.5% (chance level)
CNN, 650 nm	52.13%	68.75%

The honest takeaways were the best part of the thesis. Wavelength mattered more than the classifier: 850 nm beat 650 nm significantly (p = 0.0054) because NIR light reaches deeper into the flesh. The SVM beat the CNN: with only 110 fruits, the physics-informed features generalized better while the deep model overfit. Getting a clean, explainable 87.5% on fruit the model had never seen felt far better than a bigger number I could not trust.

CNN training curves across folds — CNN training curves: the moment it became clear the model was memorizing a small dataset.

Generalization comparison bar chart between SVM and CNN — Generalization gap between the SVM and CNN pipelines.

Running on a Raspberry Pi 4B

The winning SVM shipped to the device as a joblib model. When a scan starts, hardware PWM drives the 850 nm laser for a ~300 ms pulse, the IMX708 camera captures a lossless BMP with locked exposure, the same six features are extracted, and inference runs in about half a second. A full three-point scan (rotating the fruit between points) finishes in ≈ 1.98 s end to end, measured on the Pi itself.

The interface is a React app bundled into a single file, served by a small Flask server over Server-Sent Events, and displayed fullscreen on the prototype's 3.5" LCD. Three physical GPIO buttons control dummy scan, real scan and reset, and power. The screen reports the ripeness class, firmness category, a recommendation, and an estimated shelf-life window.