SLX SynthForge on GitHub: AI for industrial vision

SLX SynthForge, a computer vision application for electronic board inspection, is now available on GitHub. The desktop application enables operators to edit a virtual board, capture its image, and analyze defects using a locally-run AI model. Built with C++17, Qt Quick, and QML, the application relies on the proprietary Silicon LogiX SynthForge Core engine for rendering, visual analysis, and inference. Python and NumPy handle classifier training.

The reference scene, PowerBoard, features 46 editable components. Operators can manipulate components before initiating an inspection. The board model identifies nominal appearance, missing components, and changes in position or appearance. Results are categorized as PASS, DEFECT, or REVIEW, indicating accepted components, detected defects, or cases requiring review, respectively.

The vision engine, residing in the C++ engine distributed as a binary SDK, generates scene images and applies the numerical model to captures. The engine processes 640 × 480 RGB images and a defined inspection camera. Model compatibility, asset identity, and capture conditions are verified before inference.

The AI technology is a multiclass linear classifier using a softmax function. Features extracted from images are normalised and combined with weights learned during training. Softmax generates class scores, which are compared with acceptance thresholds and capture checks. Visual features, amounting to 16 per example, are used alongside three classes: nominal, missing, and changed position or appearance.

The training pipeline maintains normalisation parameters, weights, and model identifiers. The dedicated C1 and U4 component laboratory employs 24 × 24 pixel RGB crops aligned to component regions. Its model distinguishes five scenarios: nominal board, missing component, misalignment, reversed polarity, and solder bridge.

Training utilizes NumPy for numerical computation and the Adam optimiser for weight updates. The loss function is cross entropy, measuring discrepancies between classifications and example labels. Model selection relies on validation loss, with distinct roles for training, validation, and testing sets. Normalisation is fitted on training data, and test sets evaluate the model after selection. Training parameters are exported as versioned JSON files and loaded by the C++ engine for inference.

Dataset Forge generates synthetic data from the 3D scene, associating examples with simulated defect labels. This ensures experiment repeatability, case organization, and training example diversification. Capture groups are segregated into training, validation, and testing sets to minimize data leakage. The Robustness Lab alters lighting, reflections, blur, and camera tilt. Results indicate where the model maintains acceptable decisions and where a review or new capture is warranted.

The application's Qt Quick, QML, and Qt Quick Controls interface allows operators to edit boards, capture images, view inspection results, and access AI laboratories. OpenGL handles the interactive 3D preview, with a software mode for workstations that need it. C++ services coordinate processing outside the interface flow, keeping presentation, rendering, and generation tasks separate. CMake and CTest manage builds and automated checks.

The Assembly AI lab assembles development cases, prepares new examples, and compares initial models with candidates. A frozen benchmark is used for independent evaluation, ensuring both versions are assessed on the same data with the same threshold. New examples feed both training and validation. Results are documented with images, CAD references, edits, and model identity, with outcomes accessible in HTML, CSV, and JSON formats. Validation documents results and synthetic experiment conditions.