AI inspection in production
An inspection must find defects within the available cycle time without unnecessary rejects or operator interventions. Fixed vision rules are often sufficient. With greater variation, a trained model on a smart camera or in custom software can be a better fit.
Technical decision
The defect type, production variation and required explainability determine the approach. A fixed dimensional limit rarely needs AI. Variation in texture, shape or appearance can be difficult to express as separate rules.
Fixed rules
Available methods include line-to-line measurement, edge and line detection, blob analysis of connected image regions, pattern matching, counting, colour checks, OCR and code reading. These functions are not AI. With stable geometry they are often easy to explain and validate.
Model on the camera
Hikrobot smart cameras can execute supported models locally. Classical tools are also available depending on the series and software version. Sedeco can also prepare its own trained models for supported Hikrobot hardware.
Full customisation
Custom software can match the imaging, product logic, interfaces and maintenance requirements closely. Data processing, model selection, inference, error handling, updates and monitoring must then also be configured for the application.
Tool and model support differs by SKU and software version. We therefore confirm the combination during technical selection. See also the comparison of classical vision and AI.
Machine-learning pipeline
A custom model starts with a precise process decision: what must be found, what is acceptable and which defects must not be missed? We then build the dataset and model execution around that decision.
Images from different batches, positions, surfaces, lighting conditions and known defect types capture the actual variation.
With a usable CAD model, we can generate digital images using different camera viewpoints, product orientations, lighting conditions and backgrounds. This supplements situations that are difficult to collect. Real production data remains necessary to test the match with reality.
We train models, compare errors on separate validation images and optimise the model, imaging and decision thresholds for the required task.
Our production software executes the trained model, processes its output and connects it to the PLC, robot or other control system.
Annotation
Annotation means marking relevant objects, regions, classes or defects in training images. Inconsistent annotations teach the model an inconsistent task. We therefore define labels and boundary cases first with people who understand the product and process.
Sedeco has developed tools that use AI to support much of the annotation work. People review the proposals and handle uncertain cases. This makes larger datasets practical while keeping the acceptance rule in control.
Production deployment
The compute platform follows from model size, cycle time, camera count, interfaces, mounting, heat dissipation and maintenance. We test the complete chain under the conditions in which the installation must operate.
For measurement, counting, position, code and presence checks that remain reliable with fixed rules.
Local model execution on suitable Hikrobot smart cameras without a separate industrial PC.
A Sedeco-trained model, optimised and packaged for supported smart-camera hardware.
For larger models, multiple sensors or specific interfaces. We also develop and test our own hardware concepts alongside existing machine-vision hardware.
Robot vision
A 6D part pose describes its 3D position and orientation around three axes. A picking system must still determine whether the part is accessible, which grasp is possible and how the robot can reach it without a collision.
Our software combines pose estimation with grasp planning, reach checks, obstacle and collision checks and robot path planning. Recovery from an inaccessible part or failed pick is also part of the production logic.
View Robot visionAcceptance and business case
Compare investment and maintenance with manual inspection time, reinspection, rejects and rework, downtime, changeover time and operator interventions. Include data collection, annotation, model maintenance, compute hardware and integration. The outcome follows from your production data.
Required inputs: production volume, current inspection time, defect costs, line cost per hour, expected change frequency, required availability and internal review costs.
Next step
Bring representative conforming and defective parts, available images, a CAD model when available, the cycle time and the required PLC or robot interface. We can then determine whether classical tools, a smart-camera model or custom software is the smallest suitable solution.
