# Computer vision validated on images from your own production line. > Computer vision for line defect detection, counting, safety and OCR, validated on your own line images and deployed at the edge or in the cloud. Source: https://www.algofact.tech/en/services/ai-ml/computer-vision Language: en-US Updated: 2026-10-26 Languages: en-US, tr-TR ALGOFACT builds computer vision systems for manufacturing and operations sites that detect defects, count products, check safety rules and read text with OCR. We train and validate the model on real images from your line, then run it on an edge device next to the line or in the cloud, depending on latency and network needs. An operator screen, alert rules and an image archive for retraining are part of the delivery. Engineering runs from our offices in Türkiye. ## Key facts - **Tasks:** Defect detection, counting, safety, OCR - **Validation:** Real images from your own line - **Where it runs:** Edge device at the line or cloud - **Industry focus:** Manufacturing and smart manufacturing - **Model framework:** PyTorch - **Operator side:** Live screen, alerts and reject log ## When is computer vision the right solution? If a check is done by eye today, repeats all day and can be described visually, it is a strong candidate for computer vision: a surface scratch, a missing part, a wrong label, a worker entering an area without a helmet, a lot code that has to be read. A camera applies the same standard to every unit on every shift and records the result together with the image. If a defect is hard to spot even for a trained inspector, the product changes constantly or lighting and angle cannot be fixed, we recommend a feasibility study first. Real line images are collected over a period of time, we measure whether defects can be told apart in the image, and the go or no-go decision is made together on that result. That is how you avoid models that look good in the lab and fail under the lighting on the shop floor. ## Where it is used The most common tasks in manufacturing projects are quality or safety rules that can be handed over to a camera. ### Visual quality inspection Detects surface defects, scratches, stains, deformation and missing or misassembled parts. Each reject is logged with its image, and the line stops or the unit is diverted to a reject station. ### Counting and quantity checks Counts products, pallets or crates on conveyors and verifies quantities at warehouse inbound and outbound. Counts are compared with the ERP or MES record and a mismatch raises an alert. ### Workplace safety rules Real-time alerts for personal protective equipment, entry into hazardous zones and forklift and pedestrian proximity. By default the system checks the rule and does not identify people. ### OCR and label reading Reads serial numbers, lot codes, expiry dates or delivery note text and compares them with the expected value, so a mislabeled unit is caught before it ships. ### Visual measurement and alignment Measures part dimensions, alignment or fill level with a calibrated camera and flags units outside tolerance. ## Edge or cloud? Where the model runs depends on how fast the decision must be made and whether images may leave the plant. - **Edge, next to the line:** If the decision must be immediate, the line must keep working through an internet outage or images must not leave the plant, the model runs on a GPU-equipped industrial PC beside the line. Only results and sample images are sent to the central system. - **Cloud:** If images from several plants are processed together, the decision can wait a few seconds or the workload is compute-heavy, the model runs in the cloud. Bandwidth and image storage costs are estimated up front. - **Hybrid:** In most manufacturing projects the decision is made at the line while training, the image archive and reporting run centrally. New model versions are rolled out to edge devices in a controlled way, with rollback to the previous version. - **Cameras and lighting:** Most of a model’s quality comes from image quality. We settle camera position, resolution, lens and lighting during feasibility and tell you, based on measurement, whether your existing cameras are good enough. ## How the project runs ### Image collection plan We agree on which stations, products and shifts to capture. Recordings from your existing cameras can be used; any on-site work is planned separately in the discovery call. ### Defect catalog and labeling With your quality team we define what counts as a defect using sample images, then label the images against that catalog. Ambiguous cases go to your expert. ### Feasibility and proof of concept On a test set held out from real line images we measure missed defects and false alarms. Your quality team sets the acceptance threshold. ### Shadow-mode pilot The system first makes decisions without stopping the line, and we compare its calls with the operators’ calls. Thresholds are tuned on this real production data. ### Go-live and integration We go live with the operator screen, PLC or MES connection, reject signal and alert rules. Decisions and images are archived for traceability. ### Retraining loop For new products, new defect types or changed lighting, images flagged by operators are labeled and added to the next model version. ## What you receive ### Validated model A detection, classification or OCR model trained on your line’s images and measured on a held-out test set. ### Image and label archive The defect catalog, the labeled image set and the labeling rules: everything needed to retrain the model. ### Inference service A service that takes images from the camera and produces decisions on an edge device or in the cloud, with install scripts. ### Operator interface Live result screen, reject list, image review and a way to flag wrong decisions. ### Reporting and device monitoring Reports by shift, product and defect type, plus alert rules for model performance and device health. ## Questions ### Can we use our existing cameras? Often, yes. IP cameras and industrial cameras can be connected as image sources. If resolution, angle or lighting is not enough to show the defect, we measure that during feasibility and recommend additional cameras or lighting where needed. ### Can a model be trained if we have few defect samples? Yes, but the method changes. With few defect samples, anomaly detection methods that learn what a good unit looks like and flag deviations can be used. Which method is good enough becomes clear from the measurements in the feasibility stage. ### Will the line stop on a false alarm? Stopping the line is tied to a threshold agreed on from pilot results. At first the system diverts the unit or alerts the operator; automatic actions are switched on step by step as confidence grows. ### What about privacy when workers appear in the images? Safety and line monitoring cameras can process personal data under the GDPR or Türkiye’s KVKK. On the technical side we apply measures such as face blurring, short retention, restricted access and a model design that does not identify individuals. Notices and the legal assessment are handled with your counsel. ### What drives the timeline and cost? The number of stations and cameras, the variety of defect types, labeling volume, edge hardware needs, PLC or MES integration and the length of the pilot. Along with the feasibility results we present a phased scope and the hardware list in writing. ## Related - [AI, ML and Data](https://www.algofact.tech/en/services/ai-ml) - [Smart manufacturing](https://www.algofact.tech/en/industries/manufacturing) - [Data platform and MLOps](https://www.algofact.tech/en/services/ai-ml/data-platform-mlops) - [Custom AI models](https://www.algofact.tech/en/services/ai-ml/custom-ai-models) - [Discovery call](https://www.algofact.tech/en/contact) ## Sources 1. [Personal Data Protection Law (Law No. 6698)](https://www.kvkk.gov.tr/Icerik/6649/Personal-Data-Protection-Law), Personal Data Protection Authority (KVKK)