Counting Round Objects with a Smartphone: Blueberries, Seeds, Nuts and Computer Vision

The universal question: how many?

Across agriculture and food processing, thousands of decisions still rest on someone manually counting objects on a table, tray, plate, conveyor, or in the palm of a hand. Blueberries. Soybean seeds. Peas. Beans. Nuts. Pellets. Experimental samples. The counting itself is not hard — the challenge is scale. A single project can involve hundreds or thousands of images, which makes manual counting slow, repetitive, and error-prone.

Whether the objects are berries, seeds, or nuts, the underlying task is identical: count individual circular or near-circular objects from an image. Solve it once and it generalises across domains — soybean seed quality, germination studies, peanut analysis, grain research, breeding programmes, and laboratory experiments.

Where it started: blueberries

At Petiole Pro, one of our earliest counting applications came from post-harvest blueberry quality control. Researchers and industry partners wanted a faster way to quantify berry samples without spending valuable time counting fruit by hand. The workflow was collect a sample, place the berries on a tray, photograph, count, record — repeated across hundreds of images.

The counting was never the hard part — the scale was. Turning a repetitive manual chore into a few seconds of image analysis is what makes the difference at the volumes real quality control demands.

Why round objects are a good starting point

Round objects are among the easiest targets for a computer-vision system. They tend to have distinct boundaries, predictable shapes, relatively consistent sizes, and clear separation from the background. Spread them on a contrasting surface and modern image analysis can identify and count them with impressive accuracy — which is why blueberries, soybean seeds, peas, chickpeas, macadamias, hazelnuts, almonds, fertiliser pellets, and many fruits and berries all work well.

This is also why round objects make an honest starting point rather than an over-claim. Their geometry keeps the detection problem tractable, so the models can stay light enough to run on a phone. Irregular, overlapping, and biological shapes — leaves, roots, flowers — are harder, and a topic for another day.

The four-step smartphone workflow

The process is deliberately simple, so it works in a lab, a packhouse, a classroom, or a field. Everything runs from the Petiole Pro app on an ordinary smartphone.

1. Arrange the sample. Spread the objects in a single layer on a contrasting background — a tray, plate, table, or bench — and add a calibration plate if you want measurements, not just a count.
2. Capture an image. Photograph from directly above. Consistent light with minimal shadows improves accuracy; daylight is ideal.
3. Detect individual objects. Rather than treating the sample as one mass, the algorithm finds each berry or seed separately.
4. Generate measurements. Once objects are detected, the app calculates count, diameter, area, shape metrics, and distribution statistics.

Here is a step-2 capture and its step-3 result side by side — a plate of green grapes, then the same plate with every berry detected and measured:

Beyond a count: size, area, and distribution

This is where image-based counting becomes far more valuable than a total number. From a single image the app can report the number of objects, average size, minimum and maximum size, standard deviation, the full size distribution, and the percentage of objects within specific size classes. Instead of manually counting 500 blueberries, you get a complete statistical summary in seconds.

The same detection engine scales from dozens of objects to thousands. Dense seed trays are exactly where manual counting collapses and where a light on-device model keeps its footing:

Why standard deviation matters

Suppose two blueberry samples both have an average berry diameter of 16 mm. At a glance they look identical. But in Sample A most berries sit close to 16 mm — highly consistent — while in Sample B some are 10 mm and others 22 mm. Same average, very different product. For growers, packers, and buyers these are not the same thing at all, and that is what standard deviation captures.

Standard deviation measures how much variation exists within a sample. A low value means the objects are similar in size; a high value means they vary considerably. Picture a classroom: if every child is nearly the same height the standard deviation is low; if some are very short and others very tall it is high. The same principle applies to blueberries, seeds, and nuts — and it drives real decisions:

• Blueberries: retailers prefer uniform packs; a low standard deviation signals consistency.
• Soybean and chickpea seeds: uniform seed size influences planting performance and processing efficiency.
• Breeding trials: researchers hunt for varieties that produce more uniform outputs.
• Food processing: machines run better when raw materials have predictable dimensions.

Because every object is measured against the same calibrated scale, these summaries are directly comparable across samples and crops. A few real Petiole Pro results show how much the spread changes with the product:

Measured with Petiole Pro on a calibrated plate. Chickpeas, by comparison, averaged 1.10 cm in diameter with a standard deviation of just 0.09 cm — among the most uniform samples of all.

87% accuracy: precision vs practicality

Precision matters in agriculture and food processing — but how much precision is enough? Take a real example: counting 1,411 black turtle beans. The app returned 1,228 — that is 87% accuracy. Is that a problem? Not necessarily.

Why accuracy above 85% can be a win:

• Diminishing returns. Chasing the last 13% can cost more in time and resources than it is worth. 87% allows swift operations without significant compromise — and you can always count the remainder by hand.
• Nature is not uniform. Slight variation is expected and accounted for in most workflows. Staying above 85% strikes a balance between quality assurance and operational cost.
• It scales. A smartphone-based approach at this accuracy level holds up as operations grow — whether your QA team is 10 people or 100, the accuracy stays consistent across larger volumes.

Context is everything. Some industries demand higher precision; many thrive at this level. The skill is understanding your specific needs and optimising accordingly — not treating 100% as the only acceptable target.

The calibration-plate rule

Can the app count berries in a box, on a tray, or in a clump? Yes — it can count in almost any condition. But the moment you want measurements — diameter, average area, standard deviation — a calibration plate stops being a recommendation and becomes a rule.

A calibration plate acts as a measuring standard — a ruler the algorithm can trust. It tells the AI the real-world size of objects in the photo, so pixels convert into centimetres. No plate, no measurements: you will get zeros for area and diameter every time. (It is possible to measure size without a plate using a depth camera, but real-time depth processing on a phone or tablet needs extra hardware and budget.)

A free printable calibration plate ships inside the app, and sturdier acrylic plates can be purchased from the Petiole Pro store. If you are weighing up whether you need one for your workflow, we walk through the decision in detail in do I need a calibration plate for Petiole Pro?

What you can count today

The counting modules already span berries, seeds, and nuts, with dedicated quality-assurance flows in the app. Each one counts and measures; several also assess appearance and colour.

Counting is only half the story: modules such as Peanuts QA also read colour in RGB, HSV, and LAB space for appearance-based grading, and berry modules support visual quality checks that affect shelf life and marketability.

Round objects are only the beginning. Their relatively simple shape makes them an ideal entry point for automated counting and measurement, and the same thinking already extends to related tasks such as AI germination counting. Future work will tackle harder targets: irregular objects, overlapping structures, leaves, roots, and flowers.

Built for blueberries first — but the modules are extensible. If you count a specific round object that is not yet covered, a dedicated module can often be created for it.