AI Parts Counting – Choosing Weight, Vision or X-ray Counting for Thai Factories, with RFP and FAT/SAT - TOMAS TECH CO.,LTD.

Introduction to AI Parts Counting – When Someone Asks “Can an AI Camera Count This?”

You recount supplier quantities at receiving. You count screws and washers by hand at kitting. Customers complain about wrong pack quantities a few times a year. Remaining quantities on SMD reels are either counted by unwinding the reel at every stocktake or simply estimated. This article is for Japanese factories in Thailand that are considering AI parts counting in exactly these situations.

The target factories are electronic component, automotive parts and precision machinery parts plants in Chonburi, Rayong, Ayutthaya, Pathum Thani, Prachinburi and similar areas. We have in mind factory managers and people in production control, logistics and warehousing, quality assurance and IT.

Shop-floor staff turn over frequently, and the quality of hand counting varies from person to person. That is when someone inside the company asks, “Can’t we count this with an AI camera?” Yet it is hard to work out how that differs from a counting scale, how much can be left to it, and what acceptance criteria to buy against.

Here is the conclusion up front. Bringing AI into parts counting is not about “replacing human hand counting with a camera.” It is about choosing between weight counting, vision AI counting and X-ray reel counters along the following three axes.

Packaging form: loose, tray, bag, box or reel
Part characteristics: variation in piece weight, overlap, reflection, size
Cost of a mistake: whether the error happens at receiving, issuing, shipping or stocktaking

Errors in weight counting come mainly from variation in piece weight. Errors in vision-based counting come mainly from overlap, occlusion and lighting. Reels of chip components are the domain of a dedicated machine, the X-ray counter. Vision AI is most effective where parts can be laid out flat, piece weights are not uniform, and you want to keep a photo as evidence of the quantity. A pile of parts inside a box cannot, in principle, be counted from a single photo taken from above.

For that reason, PoC and acceptance criteria are not set on “average error.” They are set on the “exact quantity match rate” and the “maximum over/under count” for each item and packaging form, plus the procedure for handing back to a person when results do not match. Quantity is the entry point for traceability and inventory. The requirements should cover recording each count together with the photo, item, lot and operator, and passing it to MES, WMS and ERP.

This article does not recommend or rank any specific product. Product features and figures are presented as the manufacturers’ descriptions. The cost-benefit analysis in the second half is an estimate built on hypothetical figures for this article. The general implementation steps for vision AI, including lighting and annotation, are covered in our AI visual inspection implementation guide.

Why Thai Factories Should Revisit the Work of Counting Now

First, let us check the current state of Thai manufacturing. Figures are given with their source and scope as published. We do not add them together or apply them to any individual company.

Output expanding, employment shrinking through natural attrition: The S&P Global Thailand Manufacturing PMI, as reported by Trading Economics, was 54.3 in September 2026 (53.8 in August). This was the highest level since December 2025. The same release said employment fell slightly, citing natural attrition. Output is rising while headcount is falling.
A new employment cost contribution: According to a commentary by the law firm DFDL dated September 11, 2026, contributions to Thailand’s Employee Welfare Fund were set to start on October 1, 2026. The contribution rate is 0.25% of monthly wages each for the employer and the employee.
Extension of work permits for foreign workers: According to an article in The Star (originally from Viet Nam News) dated July 16, 2026, the Thai cabinet on July 14, 2026 approved extending the work permits of about 770,000 migrant workers from Laos, Myanmar and Vietnam. The reported aim is to secure sufficient labour for the manufacturing and service sectors.
Investment applications for automation: According to a Board of Investment of Thailand (BOI) press release dated July 23, 2026, applications for investment promotion in January-June 2026 totalled 1,299 projects worth USD 43.6 billion, up 37% year on year. Of these, applications under the “Smart and Sustainable Industry” measure (machinery upgrades, adoption of digital technology, integration of automation and robotics) numbered 132 projects worth USD 507.6 million. All of these are application-based figures.

In this environment, it becomes hard to keep absorbing counting work with more people. Hand counting requires retraining every time staff change. The way people count also differs from person to person. Revisiting the counting method is not only about saving labour, but also about making quantity quality independent of individuals.

Four Methods of Automated Parts Counting – Hand Counting, Weight Counting, Vision AI Counting and X-ray Counters

There are four broad ways to confirm the number of parts: hand counting, weight counting (counting scales), vision AI counting and X-ray counters. The choice is not about which is superior, but about which fits the packaging form and the part characteristics.

AI Parts Counting - Choosing Weight, Vision or X-ray Counting for Thai Factories, with RFP and FAT/SAT - figure 1

Method	Principle	Main sources of error	Suited packaging forms	Unsuited packaging forms and parts	Evidence	Examples (as described by each company)
Hand counting	A person counts one piece at a time	Fatigue, interruptions, person-to-person differences	Small numbers of large parts	Large quantities of small parts	Only the number written down	None
Weight counting	Total weight divided by average piece weight	Piece-weight variation, scale resolution, errors in the reference sample	Bags, boxes, trays (where the container weight is known)	Parts with non-uniform piece weight, extremely light parts	Weight and calculated count	METTLER TOLEDO, A&D GC series
Vision AI counting	Detects the individual objects in the image, or estimates density	Overlap, occlusion, lighting, reflection	Loose parts laid out flat, trays	Piles inside boxes, parts inside bags	Photo and detection result	Cognex, CountThings, in-house development (Mitsubishi Materials case)
X-ray counter	Counts by taking transmission images of reels and similar	Depends on model and part (check each company’s description)	SMD reels, ESD bags, trays	Objects outside what the dedicated machine is designed for, such as large loose parts	Image and count data	VJ Electronix, Scienscope, JFE Shoji Electronics

In short, the difference between counting scale vs vision counting is as follows. Weight counting estimates the count as “total weight / average piece weight.” Vision-based counting detects or estimates “the individual objects in the image.” The former is weak against piece-weight variation; the latter is weak against overlap, occlusion and lighting. Because their weaknesses differ, there is also room to use them in combination.

RFID and barcodes are often confused with counting. These are technologies for identifying packaging and containers that carry a label or tag. They are not technologies for counting untagged loose parts themselves. For example, reading the label on a box tells you “which item this box is and how many it is supposed to contain.” It does not tell you whether that many are actually inside. Identification and counting play different roles, so consider them in combination. How to think about the cost of using RFID for factory inventory management is covered in our article on RFID inventory costs.

How Weight Counting (Counting Scales) Works and Where It Falls Short

Counting scales are already used in many factories. Before considering vision AI, it is worth understanding where weight counting produces errors.

Scale manufacturer METTLER TOLEDO explains the following in its eBook on piece counting. First, the average piece weight is determined from a reference sample, and the count is calculated from the total weight. Errors arise from factors such as piece-weight variation. As causes of piece-weight variation, the company cites differences in production equipment and tool wear. In other words, even the same part number can differ in weight depending on the equipment or lot it was made in.

The company also explains that the referencing procedure is particularly important, because errors at the referencing stage turn into large counting errors later. It also states that an appropriate scale resolution is a prerequisite for determining the average piece weight accurately, and that scale uncertainty is a major source of counting error.

A conditional guideline is given for the reference quantity. As an example for parts whose piece-weight variation exceeds 1%, the company explains that increasing the reference quantity from 10 to 30 pieces can halve the counting error. This is an example, not a figure that applies to all parts. The company also states that using a sample of 30 or more pieces can improve accuracy.

As an example of a feature, the company introduces “Automatic Reference Optimization.” According to its description, you first count 10 pieces by hand to determine the average piece weight, and the scale then automatically recalculates the average piece weight using the additional pieces placed on it. As a configuration example, it also shows a two-scale setup combining a 0.1 mg reference scale with a 0.6 kg to 6 kg counting scale. The company explains that a two-scale setup can cover a range from 1 mg to 3,000 kg. Example applications listed are receiving, warehousing, production, and logistics and shipping. A configuration in which item information is called up from a barcode scanner at receiving and sent to ERP or MES is also given as an example.

Let us also look at an example from a Japanese manufacturer. A&D announced its “GC series” counting scales for stocktaking and bagging work on January 26, 2023, and released them on February 1, 2023. According to the company, they have the following features.

ACAI: an automatic accuracy improvement function that counts while correcting the registered piece-weight value
AIS: a function that automatically searches the multiple piece weights registered in internal memory for the one closest to the item being weighed (the company states that it is patented)
Memory for 1000 items, three-colour comparator lights (LO, OK, HI), connection to devices such as barcode readers via an optional multi-interface, and piece-weight registration using the PC software “WinCT-Counting”

The company’s product page states that the standard sample size is 10 pieces, and that 5, 25, 50, 100 or any number of pieces can be set. The minimum weighable weight and minimum piece weight are listed as 0.1g and 0.005g for the GC-3K, 0.2g and 0.01g for the GC-6K, 0.4g and 0.02g for the GC-15K, and 1g and 0.05g for the GC-30K. The standard list price (excluding tax, in Japan) is JPY 89,000 for the GC-3K and GC-6K, and JPY 92,000 for the GC-15K and GC-30K. We could not confirm prices in Thailand.

To sum up, the weakness of weight counting is piece-weight variation, and manufacturers provide features to correct for it. For parts with uniform piece weights and packaging forms whose container weight is known, weight counting remains a strong option today. In some cases, simply reviewing the referencing procedure and resolution of your current scales can bring improvement before you introduce vision AI. How to bring scale readings into your systems is covered in our article on scale data integration.

How Vision-Based Parts Counting Works – Object Detection and Density Estimation

AI that counts from images falls into two broad families. A survey paper on counting techniques revised in February 2026 (Ciampi et al., arXiv) organises them as follows.

Detection-based: an object detector finds each individual object, and those are counted
Regression-based: the model learns a mapping from image features to a density map, and the count is estimated as its sum

The same survey writes that detection-based methods are weak against occlusion. For that reason, many of the “class-agnostic counting” methods described below use the density map approach. A density map is created by placing a point at the position of each object in the training image and placing a blur around it.

On the factory floor, this difference shows up as follows. Detection-based methods tell you the position of each object, so it is easy to check on screen “which ones were counted.” Density-based methods can estimate the count even with some overlap, but they do not readily show the position of each individual object. Which suits you depends on the packaging form and on how operators want to check results.

Another important concept is class-agnostic counting. According to the survey, this is a technique for counting types of objects not seen during training, using a small number of examples (few-shot). The survey classifies it into three types: methods that use exemplar boxes, methods that automatically find the objects to count, and methods that specify the target by text.

A representative research example is FamNet, presented at CVPR 2021, and the FSC-147 dataset created for it (Ranjan et al.). When you draw boxes around a few examples in an image, it predicts a density map for all objects of the same type and counts them. FSC-147 consists of 6135 photos of general objects across 147 categories. These are not photos of factory parts. Note that error values from research do not indicate the accuracy you would get when counting parts in a factory.

The practical meaning is that there are methods that can start counting once you show a few examples, without collecting a large number of training images for each item. Whether you can then use them to make pass/fail decisions at receiving or shipping is a separate question. You need to verify this by running in parallel on your own items and packaging forms.

What Vision AI Counting Struggles With – Overlap, Piles, Reflection and Transparency

Let us confirm what vision AI struggles with, based on research and product descriptions.

“Counting Machine Parts” (Arockiaraj et al., arXiv), submitted in May 2026, deals with counting washer parts for machinery. As the main difficulties, this report lists overlapping objects, differences in object scale, occlusion and variation in lighting conditions. We use only this problem setting here.

“Automated Counting of Stacked Objects in Industrial Inspection” (Dumery et al., arXiv), submitted in March 2026, deals with manufactured parts stacked inside containers. The starting point of this research is that most of the stacked parts are hidden from view. It therefore proposes a method that counts by estimating 3D shape and occupancy from images taken from multiple viewpoints. In other words, a pile inside a box cannot, in principle, be counted from a single photo taken from above, because some objects are not visible.

From these two studies, the following can be said about the shop floor.

Overlap and contact: When parts overlap or touch each other, the boundary of each piece becomes hard to see. In many cases, spreading the parts out flat is a prerequisite
Piles: Piles inside boxes or bags cannot be counted from a single photo taken from above. Either spread them out and count, or combine with weight counting
Differences in size: When large and small parts are mixed, counting errors become more likely
Variation in lighting: When the appearance changes with light from windows or the time of day, results fluctuate

On reflective parts, there is Cognex’s description of the AI counting tool for In-Sight SnAPP announced on August 6, 2024. The company explains that it can count objects of similar size, reflective or distorted parts, and objects with different contrasts and shapes. This is the manufacturer’s claim. Put the other way round, it also means that reflection and distortion are generally difficult to count. Please verify with your own parts.

Transparent parts and light-transmitting plastic parts are also difficult targets. Backlighting, which brings out the outline like a silhouette, can help separate parts in some cases. However, the effect varies by part and equipment, so we do not estimate it numerically. How to choose lighting is covered in our article on AI visual inspection lighting selection, and how to decide camera resolution and field of view is covered in our article on industrial camera selection.

X-ray Component Counters – Chip Component Inventory Is the Domain of Dedicated Machines

SMD chip components are stored wound on reels. You cannot tell the remaining quantity on a reel by looking at it. Counting by unwinding it by hand takes time. In this domain, dedicated machines that count by taking X-ray transmission images of reels are already in use. It is easier to organise your thinking if you treat vision AI as an option for everything else: loose parts, trays and the contents of boxes.

All of the following are descriptions by the manufacturers or distributors.

VJ Electronix XQuik III: The product page of distributor DKSH explains that it meets the industry accuracy standard of 99.8%. The page does not state for which parts or under which conditions this figure applies. The same page also lists specifications such as barcode reading and count results in under 10 seconds, automatic barcode scanning, an imaging resolution of 130μm, and 60kV X-ray
VJ Electronix XQuik V: According to a Circuits Assembly article dated January 6, 2026, the company explains that it features X-ray component counting and inventory management, an automatically loading drawer, and a barcode camera that links reel identification with count data, reducing manual entry errors
Scienscope AXI5100c: According to an article dated February 16, 2016, it is a fully automatic inline X-ray component counter that can count parts as small as 03015 size. The company explains that it can handle four 7-inch reels or one 15-inch reel, and can also count parts in ESD bags, sticks and trays
Scienscope AXC-800 III: According to an article dated August 21, 2019, the company explains that it counts four 7-inch reels within 23 seconds and one large 13-15-inch reel within 16 seconds. A built-in barcode scanner and automatic label printing are also described
JFE Shoji Electronics iNsight-2000: According to the company’s page, it can measure up to four 7-inch reels simultaneously and can be linked with ERP and MES. The company explains that it can count without opening aluminium moisture barrier packaging (ESD, MSD)

We could not confirm published prices for any of these. JFE Shoji Electronics lists the price as “on inquiry.”

The reasons for considering an X-ray component counter are not limited to labour hours. Other reasons include wanting to count without opening moisture barrier packaging (MSD control), wanting to enter reel identification and remaining quantities accurately into the inventory system, and customers requiring control of component inventory. As the estimate later shows, in factories with few reels it may be difficult to recover the investment through labour savings alone.

Examples of Products and Services, and an In-House Development Example

Here are concrete examples of vision AI counting from three angles: a product, an app and in-house development. All are descriptions by the manufacturers or from the case studies, not recommendations.

Cognex (counting tool for vision sensors): On August 6, 2024, the company announced an AI counting tool for its In-Sight SnAPP vision sensor (confirmed in the company release as distributed). The company explains that it can be used to automate assembly verification and quantity checks. What it can count was covered in the previous section. It is a vision sensor fixed to the line. We could not confirm a published price.

CountThings (smartphone and PC app): An app provided by Dynamic Ventures that detects and counts objects from photos and videos. The company explains that it runs on iOS, Android and Windows. The objects to count are specified by a “counting template.” According to the company’s Japanese-language page, it supports customers by analysing their sample images to optimise existing templates or create new ones, and in most cases this optimisation service is free. The company explains that the app counts offline on the device and protects data privacy.

On accuracy, the company’s site carries a statement to the effect of “please contact us if you are not getting perfect accuracy.” This wording assumes that accuracy varies with the object and the template. It does not mean that any object can be counted perfectly. Example uses listed include tablets (pills), pipes, logs, plasterboard and lumber.

The prices listed on the company’s site (as viewed in October 2026) are as follows. Monthly is USD 120 per device, and annual is USD 1,200 per device. 24-hour licences range from USD 24 for one to USD 150 for ten, and pay-per-use is USD 80 per 100 counts. “CountThings from Video,” which counts from video, is USD 600 per device per month or USD 2,400 per year.

Mitsubishi Materials (in-house development case): This is a case study published on June 24, 2025 by cloud provider iret (cloudpack). To make parts counting work in its factory more efficient, Mitsubishi Materials developed an AI image analysis app using YOLO, an object detection method. It is built on AWS (Amazon SageMaker), with a React front end. The targets include drills for grinding and insert parts.

As an effect, the case study states a reduction in workload of about 3 person-days per month. Training is said to have used 30-40 images per part type. Accuracy is not given as a number; the description is that the balance between accuracy and response speed was highly rated. This case concerns a specific factory and specific parts of the company in Japan, not a site in Thailand. Development costs are not stated either.

The three examples take completely different forms: a sensor fixed to a process, an app carried by the operator, and a system built in-house. How to choose among them is organised in the later section on “Choosing a configuration.”

Where to Start – Comparing Quantity Verification Automation Across Receiving, Issuing, Shipping and Stocktaking

There are four main situations in which parts are counted in a factory: receiving, issuing (kitting), shipment packing and stocktaking. The cost of a mistake and the packaging form differ by situation, so the suitable method also changes.

Situation	Cost of a mistake	Common packaging forms	Candidate methods	Need for evidence
Receiving	Accepting a short delivery as a full quantity makes inventory appear higher than it really is, leading to shortages later	Bags, boxes, trays, reels	Weight counting, vision AI counting (where parts can be spread out), X-ray (reels)	Photos help in negotiations with suppliers
Issuing and kitting	Issuing too many builds up surplus on the floor; issuing too few stops the line	Loose parts, small bags, trays	Weight counting, vision AI counting	Records of operator and lot make it possible to trace causes
Shipment packing	Packing too few leads to shortage complaints from customers; packing too many means giving parts away free	Boxes, bags, trays	Weight counting, vision AI counting (where parts are laid out before packing)	Photos help in explaining to customers
Stocktaking	Book and physical inventory diverge, disrupting ordering and production planning	Boxes, bags, reels	Weight counting, X-ray (reels), vision AI counting	Records are needed to investigate discrepancies

What this table shows is that the same “one piece too many” means different things in different situations. Counting too many at receiving and packing too few at shipping differ in who bears the loss and in how large the loss is. When choosing where to start, first take stock of where over/under counts are occurring now and how large those losses are.

Building the system for each situation is covered in detail in separate articles. For receiving, see our article on receiving inspection systems. For issuing, see our article on production material issue control and our article on automating line-side parts supply. For shipping, see our article on shipping and receiving inspection systems and our article on preventing mis-shipments.

If you start with vision AI, situations where parts can be laid out flat are candidates. Examples are a kitting process where small parts are spread on a tray before being issued, or a process where parts are laid out on a tray for checking before shipment. Where you want to count parts still in bags or boxes, weight counting is often more suitable.

Training Data and Labels – How to Manage the Item Master and Images

The time-consuming part of vision AI counting is not the AI itself but managing images and master data for each item. In factories with hundreds or thousands of items, this becomes the core of operations.

There are three ways to think about it.

Registered images per item: Link each item in the item master to a representative photo and to how the object to be counted appears (front and back, orientation). Make sure you can tell which images are outdated when a part number changes or when the appearance changes because of a supplier change
Methods using a few boxed examples and pre-trained models: With products and apps that use the class-agnostic counting approach, you can start counting by drawing boxes around a few examples on screen. While you do not need to collect a large number of training images for each item, accuracy for each item and packaging form is verified by running in parallel
Methods that train per item: The Mitsubishi Materials case states that the model was trained with 30-40 images per part type. This is a figure for that company’s specific parts and conditions. The number required varies with how the parts look and with the packaging form, so we do not treat it as a general guideline

Also decide at the outset the procedure for adding items. Who takes the photos? Who draws boxes around the examples, or applies the labels? Who reviews the results of the parallel run and approves use? Items not yet approved are handled by hand counting or weight counting. Without such a procedure, items get added on the shop floor at will and are used without any accuracy check.

The quality of labelling also matters. If the correct count for a photo is wrong, both evaluation and training go astray. Measures include having two people count the correct answer and cross-check, and also checking it against the weight counting result. How to run labelling is covered in our article on the visual inspection annotation workflow.

Setting Acceptance Criteria for an AI Counting System – “Exact Match Rate” and “Maximum Over/Under Count,” Not Average Error

A metric often seen in PoCs and acceptance tests is “average error.” In parts counting, however, average error alone is not enough to judge. For example, a case that is exact 99 times out of 100 but is far off once, and a case that is slightly off every time, can have similar average errors. What causes problems on the shop floor is the single large miss.

So for each item and packaging form, look at the following metrics.

AI Parts Counting - Choosing Weight, Vision or X-ray Counting for Thai Factories, with RFP and FAT/SAT - figure 2

Metric	Definition	Data	How to judge
Exact match rate	Number of times the AI count exactly matched the correct count / total number of counts	Parallel-run records for each item x packaging form	Set the target value yourself (example value)
Maximum over/under count	The largest number of pieces by which a single count was above or below the correct count	Same as above	Decide the maximum acceptable over/under from the cost of a mistake in that situation
Number of overcounts and undercounts	Count separately the times the result was above and below the correct count	Same as above	The direction of loss differs by situation, so judge them separately
Hand-back rate	Number of counts sent to human recount due to low confidence or mismatch / total number of counts	Operations log	Too many and labour hours do not fall. Too few and misses increase
Results after hand-back	Difference between the human recount and the AI count	Recount records	Use as material to review whether the hand-back conditions are appropriate

There is no industry standard for target values. Decide them from your own situations and losses, start the PoC with example values, and review them based on the results.

Especially important is the asymmetry between overcounting and undercounting errors. If the AI overcounts at receiving, you accept a delivery as a full quantity when it is actually short, which leads to shortages later. If the AI overcounts at shipping, you have in fact packed too few, and the customer complains of a shortage. Conversely, undercounting leads either to packing extra and giving parts away, or to more recounting work. Which error to suppress more strongly differs by situation.

When you cannot be confident in the AI’s result, or when it does not match another method such as weight counting, hand it back to a person for recounting. Only when this “hand-back procedure” works can AI counting be used in operations. The idea of designing for over-detection and misses together is the same in visual inspection. For details, see our article on controlling false rejects in AI visual inspection.

Also pay attention to the period of the evaluation data. If you evaluate only with images from the same day, the same operator and the same lot as those used in the PoC, results tend to look good. Evaluate over a period that includes different days, different shifts and different lots.

Records and System Integration – Sending Photo, Item, Lot and Operator to MES/WMS/ERP

Quantity is the entry point for traceability and inventory. If counting results are only written on paper, introducing AI delivers only half the benefit. Keep the following information as a single record and pass it to your systems.

Date, time and place of counting (receiving, issuing, shipping or stocktaking)
Item, lot, and supplier or ship-to destination
Counting method (vision AI, weight, X-ray, hand counting) and the resulting count
The photo used for counting (for vision AI or X-ray)
Operator, and if handed back to a person, the reason and the recount result

Barcodes are useful for calling up items. METTLER TOLEDO gives as an example a configuration in which item information is called up from a barcode scanner at receiving and the counting result is sent to ERP or MES. The same idea can be used with vision AI counting. If you first read the barcode on the box or the parts tag, call up the template for that item and then count, you can prevent selecting the wrong item.

Saving photos pays off later. When you tell a supplier that a delivery was short, or when a customer asks about a wrong pack quantity, photos and records let you show how many there were at that point. Decide the retention period and storage location in line with customer requirements and internal rules.

The systems to integrate with differ by factory. Is inventory managed in WMS or ERP? Are lots tracked in MES? Who approves counting results before they are reflected in inventory? Sort out these points before introduction. Handling of discrepancies found at stocktaking is touched on in the next section.

Stocktaking and Inventory Discrepancies – “Consistency of Counting Method” Matters More Than Counting Speed

At stocktaking, attention tends to go to counting speed. To reduce the causes of inventory discrepancies, however, consistency in how you count matters more. If the same item is counted by hand by some people, weighed on a scale by others, and taken on trust from the box label by others still, the causes of discrepancies become impossible to identify.

Mismatches in inventory records are not unusual. In a 2008 study by DeHoratius and Raman that examined about 370,000 inventory records across 37 stores of a single US retailer, 65% of inventory records did not match actual inventory. This is a retail study, and the figure does not apply directly to parts inventory in manufacturing. However, the same study reported that inventory audits were a factor that mitigated the mismatches. This suggests there is value in putting a counting system in order.

When deciding counting methods for stocktaking, decide for each item that “this item is counted with this method” and hold that in the master. For example, reels by X-ray, small parts with uniform piece weights by weight, and loose parts that can be laid out by vision AI. If you record the method, the result and the photo, then when a discrepancy appears you can isolate whether it is a counting problem or a problem with the receipt and issue records.

How to shorten stocktaking time is covered in our article on reducing inventory count time, how to investigate the causes of discrepancies in our article on the causes of and countermeasures for inventory discrepancies, and how to run physical stocktaking in our article on stocktaking efficiency.

Choosing a Configuration – Smartphone App, Fixed Station or Edge/On-Premises

There are three broad configurations for vision AI counting.

Configuration	Suited situations	Advantages	Cautions
Smartphone or tablet app	Many items, counting locations scattered. You want to try it first	Quick to introduce. Portable	Shooting distance, angle and lighting change every time. Device management is needed
Fixed station (camera, lighting, PC)	Counting at set places such as the receiving bench, kitting bench or packing bench	Shooting conditions are consistent and results tend to be stable. Easy to connect to records	Requires installation space and cost. Parts must be brought to the bench
Processing on an edge or on-premises server	You do not want photos to leave the factory. The network is unstable	Data stays on site. Fast response	You must run equipment maintenance and model updates yourself

The first thing to consider is data leaving the site. Confirm whether photos of parts and item information may be sent to the cloud. CountThings explains that its app counts offline on the device. Whether a configuration can be used without sending data to the cloud is a point of comparison when evaluating products.

Also check the factory network. If Wi-Fi is weak at the back of the warehouse or in the packing area, results cannot be sent to the system. In some cases you need a mechanism that counts offline and sends the results in batches later. How to think about edge processing configurations is covered in our article on edge AI deployment in factories.

Work continues after operation starts: adding items, replacing images for items whose appearance has changed, and monitoring for declines in accuracy. Accuracy can drop with changes such as a supplier change altering the colour of a part, or replacing the lighting. Measure the match rate confirmed in the parallel run periodically during operation as well. For how to monitor and retrain models, see our article on factory AI model monitoring.

Requirements to Write in the RFP and FAT/SAT Acceptance Criteria

Before asking vendors for quotes, put in writing what you want to count. The table below is an example of requirements to write in an RFP (request for proposal). Decide numerical targets yourself, starting from example values.

Item	What to write
Target items and packaging forms	List of items, and whether each is counted laid flat, on a tray, in a bag, in a box or on a reel
Number of items	Number of items at start, and the expected number of items added per year
Maximum count per session	The maximum number of pieces handled in one count, and the range of part sizes
Allowed placement	Flat placement only, trays allowed, or whether counting in the bag is required
Acceptance criteria	Exact match rate and maximum over/under count for each item x packaging form. Period and number of cases of the evaluation data
Procedure on mismatch	The on-screen flow for handing back to a person on low confidence or mismatch, and recording of the recount
Image storage	Images to store, retention period, storage location, and how to search
Master integration	Synchronisation with the item master, calling up items by barcode, sending results to MES, WMS and ERP
Permissions and change history	Who can change templates and models, and whether changes are recorded
Item addition procedure and cost classification	Whether you can add items yourself or must ask the vendor. How cost per item is handled
Training and maintenance	Training for operators and administrators, a support contact, and the scope of maintenance
Quote breakdown	Ask for hardware, software licences, item registration work and maintenance to be quoted separately

The reason for asking for a breakdown is to make comparison easier. Even if the hardware is cheap, the work cost of item registration may be high. The total also changes depending on whether licences are per device, per site or per number of counts.

Think of acceptance testing in two stages.

FAT (factory acceptance test, before shipment): On the vendor’s side, reproduce counting using actual past samples from your own factory. Confirm whether the RFP acceptance criteria are met for each target item and packaging form
SAT (site acceptance test): At your own factory, use it in the same flow as actual operations. Run it in parallel with human hand counting or weight counting and reconcile the results. Check the exact match rate, maximum over/under count and hand-back rate for each item and packaging form

Stop hand counting starting with the items that met the criteria at SAT. For items that do not, decide whether to change how they are placed or keep them on weight counting. There is no need to switch all items at once.

Quantity checks may also be used as poka-yoke in assembly processes, for example to confirm that the specified number of screws was used. In that case, the purpose becomes preventing missed work steps rather than counting. That way of thinking is covered in our article on poka-yoke.

Model Estimate of Costs and Benefits (Hypothetical Figures)

The figures from here on are all hypothetical figures for this article. Official product prices could not be confirmed for many products, so costs are also placeholders. Please read this as a framework for running your own estimate.

The assumed factory is a Japanese electronic component or automotive parts factory in Thailand. Small parts are counted at three places: receiving, kitting and shipment packing. It runs two shifts, 250 days a year.

Current baseline (per year)

Item	Amount	Breakdown
Labour hours for hand counting and recounting at receiving, kitting and shipment packing	720,000 THB	6,000 hours x 120 THB
Handling caused by quantity errors (recounting, investigating inventory discrepancies, dealing with over/under shipments to customers)	600,000 THB	Placeholder
Labour hours for checking remaining quantities on SMD reels at stocktaking and receiving	300,000 THB	2,500 hours x 120 THB

Configuration A: three vision AI counting stations

Includes cameras, lighting, PCs, software, integration with the item master and MES, and item registration. Initial cost is set at 1,500,000 THB, and annual operating cost (licences, item additions, maintenance) at 180,000 THB.

Benefit item (assumption)	Calculation	Amount
50% reduction in hand-counting labour hours	720,000 x 0.5	360,000 THB/year
30% reduction in handling caused by quantity errors	600,000 x 0.3	180,000 THB/year
Total benefit	360,000 + 180,000	540,000 THB/year
Annual net benefit	540,000 – 180,000	360,000 THB/year
Payback period	1,500,000 / 360,000	About 4.2 years

Configuration B: Configuration A plus one X-ray reel counter

Additional initial cost is set at 2,000,000 THB, and additional annual operating cost at 100,000 THB.

Additional benefit item (assumption)	Calculation	Amount
80% reduction in labour hours for checking remaining reel quantities	300,000 x 0.8	240,000 THB/year
Annual net benefit of the addition	240,000 – 100,000	140,000 THB/year
Payback period of the addition	2,000,000 / 140,000	About 14.3 years

Sensitivity: Configuration A with the hand-counting labour baseline doubled

Suppose the hand-counting baseline was 1,440,000 THB/year. A 50% reduction in hand counting is 720,000 THB/year. Total benefit is 720,000 + 180,000 = 900,000 THB/year, net benefit is 900,000 – 180,000 = 720,000 THB/year, and the payback period is 1,500,000 / 720,000, or about 2.1 years.

AI Parts Counting - Choosing Weight, Vision or X-ray Counting for Thai Factories, with RFP and FAT/SAT - figure 3

What can be read from this estimate is as follows.

The payback period of Configuration A changes greatly depending on how many hand-counting labour hours there are. If the baseline doubles, the payback period shrinks from about 4.2 years to about 2.1 years. Measuring your current hand-counting time is the starting point of the estimate
The addition in Configuration B takes about 14.3 years to pay back on labour savings alone. This means that in factories with few reels, an X-ray counter is hard to pay back on labour savings alone. Add the number of reels, stocktaking frequency, MSD control (whether you want to count without opening packaging) and customer requirements to the decision
The “30% reduction in handling caused by quantity errors” assumes that the procedure for handing AI counting errors back to a person is working. Do not count this benefit until you have been able to confirm the match rate through a parallel run

Frequently Asked Questions

What is an AI counting system, and how does it differ from a counting scale?

AI counting is a method that counts the number of parts by detecting the parts in an image taken by a camera, or by estimating their density. A counting scale obtains the count by dividing the total weight by the average piece weight. The difference is that scales are weak against piece-weight variation, while vision AI is weak against overlap, occlusion and lighting. Another difference is that vision AI can keep the photo of what was counted as evidence.

How accurate is vision-based parts counting?

It varies greatly with the target parts and packaging form, so it cannot be answered with a general figure. Errors shown in research are measured under conditions such as photos of general objects, and do not apply directly to your own parts. Judge by running in parallel with human hand counting or weight counting on your own items and packaging forms, and measuring the exact match rate and maximum over/under count for each item.

Can parts be counted while still in bags or boxes?

With vision AI, counting a pile in a box from a single photo taken from above is difficult in principle, because some parts are not visible. Research has proposed methods that count from images taken from multiple viewpoints, but on the shop floor, spreading the parts out to count them or using weight counting are the realistic options. Chip components in reels or ESD bags are targets for X-ray counters.

Is a smartphone app enough for automated parts counting?

At a stage where you have many items and want to try it first, a smartphone app is a strong option. However, because shooting distance and lighting change every time, results can fluctuate. If you will keep formal quantity records at receiving or shipping, consider whether you need a fixed station with consistent shooting conditions and integration with the item master and your systems.

What is an X-ray component counter used for?

It is mainly used to count the remaining quantity on reels wound with SMD chip components. Manufacturers explain that it can count without unwinding by hand, and depending on the model, without opening moisture barrier packaging. Various companies also describe configurations that link reel identification with count data via barcode and send it to ERP or MES.

How many training images are needed?

It depends on the method. With methods that can start counting once you box a few examples, you do not need to collect a large number of images for each item. As an example of a method that trains per item, the Mitsubishi Materials case states 30-40 images per part type. This is a figure for that company’s specific parts and conditions, not a general guideline.

How long does implementation take?

It depends on the number of items and the variety of packaging forms. It will be shorter if you have few items and start with parts that can be laid out flat, and longer if you have many items and diverse packaging forms. A realistic approach is to expand the scope while confirming acceptance for each item and packaging form at each stage of PoC, FAT and SAT.

Summary

Bringing AI into parts counting is about choosing between weight counting, vision AI counting and X-ray reel counters according to three factors: packaging form, part characteristics and the cost of a mistake
Errors in weight counting come mainly from piece-weight variation; errors in vision-based counting come mainly from overlap, occlusion and lighting
A pile inside a box cannot, in principle, be counted from a single photo taken from above. Vision AI is most effective where parts can be laid out flat, piece weights are not uniform, and you want to keep photo evidence
Reels of chip components are the domain of a dedicated machine, the X-ray counter. In factories with few reels it is hard to pay back on labour savings alone, so the decision should also include MSD control and customer requirements
Acceptance criteria are set not on average error but on the exact match rate and maximum over/under count for each item and packaging form, with mismatches and low-confidence results handed back to a person
Requirements should cover recording counting results together with the photo, item, lot and operator, and passing them to MES, WMS and ERP
In the RFP, ask for quotes broken down into hardware, licences, item registration and maintenance, and at SAT run in parallel and reconcile the results

TOMAS TECH is available for consultation even at the stage of taking stock of which items and packaging forms to start with, and where over/under counts are occurring with your current counting methods. Please contact us via this form.

AI Parts Counting – Choosing Weight, Vision or X-ray Counting for Thai Factories, with RFP and FAT/SAT