Can Cameras Catch Counterfeits? Real-World Q&A

18/09/2026

People often ask: "If we install cameras, will AI automatically catch every fake bottle of water or swapped package?" That expectation sounds logical. But in reality, when sitting down with operations teams, we often face flickering fluorescent lights, goods dented during transport, and old tape residue. Computers do not "see" like humans. They only read pixels. If you expect a magic scanner to solve all fraud issues from warehouse intake to the moment a customer pays, you will be disappointed. Here, I will candidly share what has happened in retail Computer Vision projects I have been involved in, from major food chains in Vietnam to distribution factories in Thailand.

Is image-based counterfeit detection absolutely accurate?

No. This is the most straightforward answer you will hear. Computer Vision has high accuracy, but it depends entirely on the quality of the training data. If you only teach the machine 500 images of genuine water bottles, it will not know how to detect fake bottles with sharper printed labels, or counterfeit labels with blur levels equivalent to the real ones. In a project for the lubricant industry, we discovered that over 30% of recognition errors came from warranty labels being misaligned by just a few millimeters. Humans can spot that anomaly, but the machine needs to be "taught" that this misalignment is a warning sign. Therefore, do not trust promises of "99% accuracy" if they do not show you their dataset. Start with a small sample set, test in the field, and accept that the system will need continuous tuning in the first few months.

How do you handle dented goods in the warehouse?

This is the hardest problem. Goods in the warehouse are never perfectly flat on the conveyor belt. They are stacked, bumped, and their packaging is deformed. If you use standard face or logo recognition algorithms, the error rate will be very high. The way we usually handle this is by separating the process. At the intake stage, instead of requiring the machine to identify the product, we focus on checking the integrity of the packaging. The system looks for tears, stains, or structural surface differences compared to the standard sample. We implemented this method with an instant noodle company. Instead of trying to read a barcode that is folded in half, the camera compares the overall shape of the noodle pack. If the curvature does not match the reference sample, the system moves that package to a manual inspection tray. It may seem less efficient than full automation, but it minimizes the risk of missing counterfeit or defective goods that humans easily overlook when working at high speed. This is a reasonable trade-off: accepting some manual work in exchange for brand safety.

Is the hardware cost for product safety expensive?

Yes, but not as expensive as you think if you know how to choose. A common mistake is buying 4K resolution cameras for every location. You do not need 4K to see a small label on a water bottle at a distance of 30cm. You need a telephoto lens and enough resolution to clearly read the font printed on the label. In a project at a factory in the Philippines, we saved nearly half the hardware cost by only installing high-resolution cameras at critical inspection points (outlet, labeling area) and using standard cameras for general monitoring positions. However, do not cut back too much on lighting. Light is a vital factor. A cheap camera illuminated by specialized LED lights will deliver results twice as good as an expensive camera placed in the dark or mixed lighting. Invest in a uniform lighting system and avoid shadows. This is something many software providers overlook, but deployment engineers like me consider it the top priority.

How much training data do you need to start?

Less than you think, but quality is more important than quantity. You do not need a million images to start. With transfer learning techniques, we can train a model to recognize a new product line with just a few hundred sample images, provided those images capture various lighting conditions and viewing angles. The big issue is not a lack of genuine product images, but a lack of counterfeit images. How do you get counterfeit images to train the machine? We often use data synthesis techniques: creating simulated variants of counterfeits (blurred labels, wrong colors, misalignment) based on reports of common counterfeiting methods in the region. For example, if the Thai market frequently encounters photocopied labels, we will create a series of images simulating photocopy print quality to include in the training set. This method helps the model learn "non-genuine" characteristics without actually possessing counterfeit goods. This is a crucial step to enhance product safety in the early deployment phase.

Does image processing speed affect the conveyor belt?

Yes, and this is a physical limit you cannot ignore. A camera taking a photo and sending it to a server for processing can take 200-500ms. If your conveyor belt runs at high speed, this number can cause bottlenecks. In large factories, we often deploy edge computing. This means placing a powerful computer right next to the camera, or using AI-integrated circuit boards. This machine processes the image on the spot, only sending the result "OK" or "Error" to the central system. This reduces latency to under 50ms, fast enough for the system to push defective goods off the conveyor belt before they travel far. If you are considering deployment in multinational retail chains, carefully check your internal network bandwidth. Sometimes, upgrading the control room server is more important than buying more cameras. A slow network system will turn an advanced AI solution into an operational bottleneck.

Can AI completely replace human goods inspectors?

Absolutely not. Maintain this humility. AI is a support tool, not a supervisor. In practice, we always design systems so that AI points out suspicions, and humans make the final judgment for difficult cases. Humans have the ability to reason in context, which machines do not yet have. For example, a stain on packaging could be due to counterfeit goods, or rainwater during transport. The machine will warn "abnormal stain detected," but the inspector will look at the batch history to decide whether to keep or discard it. Completely removing humans is not only a technical risk but also a legal liability risk. When a product harms a customer, you need a human to be accountable for that decision. AI can be wrong, but humans cannot (legally). Therefore, the most effective model today is collaboration: AI handles repetitive, boring, and precise tasks; humans handle analysis, exceptions, and accountability.

Where to start deploying retail Computer Vision?

Start from a specific pain point; do not try to solve everything at once. Do not think about installing cameras for the entire warehouse and every checkout counter from the start. Choose a location with high risk and high value. For example, if you are concerned about counterfeits at the checkout, try installing it in a pilot store with high customer traffic. Run the system for 4-6 weeks. Collect data. See what the false positive rate is. If 10 alerts result in 8 false alarms, the system will annoy staff, and they will ignore it. That is a failure. You need to fine-tune the algorithm or change the camera position to reduce this rate to under 2-3%. Once the system is stable at one point, scale it out. We did this with Masan and Meat Deli, starting from a few factories, then expanding to the entire chain. Patience in the early stage will help you build a solid foundation. Do not rush. AI in product safety is a long-term game, requiring continuous adjustments for each season, each batch, and each new type of fraud appearing in the market.

AIVISION helps enterprises turn AI into working systems. Explore our enterprise AI solutions, read more on the AIVISION blog, or talk to our team about your own use case.