Visual defect detection in factory automation

Training data will be images collected from a camera that closely matches the one used on the factory line. Lighting should be consistent with the factory conditions, including any additional illuminators like ring lights. The camera lens can also impact image quality, so it should match the lens (or lenses, if multiple options will be available) used in the final system.

Ground truth for the dataset can be labels as bounding boxes over defective parts (better for small, simple parts like screws) or image masks that isolate pixels on specific regions showing defects (better for large, complex objects like PCBs). Select the model type before labeling data, as the model architecture may influence the type of annotations required.

Models may be trained with CCStudio™ Edge AI Studio or edgeai-modelmaker for TI-supported neural networks. More advanced developers may approach this task using autoencoder models for anomaly detection.

Select the best model type for your defects:

• Object detection models (like YOLOX): Best for spotting distinct, separate defects such as cracks, missing components, or other large defects.
• Semantic segmentation models (like DeepLabv3): Ideal for detecting surface-level issues like discoloration or texture problems.

Balance resolution and speed:

Choose a resolution high enough to catch important small defects, but not so detailed that it slows down your inspection process. Most systems improve efficiency by downscaling the original captured images to reduce the amount of data to process.

Consider anomaly detection for unique cases:

For situations where defects are unpredictable or hard to define, anomaly detection models (using autoencoders) can offer a simpler approach. These models learn what "normal" products look like and flag anything different as a potential defect—no need to train on defective examples. However, these types of models are not supported in Edge AI Studio or edgeai-modelmaker, so developers must implement an autoencoder neural network themselves and export to a supported format like ONNX. 

Model deployment requires the model to be compiled beforehand for the target hardware accelerator. With tools like Edge AI Studio and edgeai-modelmaker, compilation is automatic. Otherwise, compiling models will require a separate step through software packages like edgeai-tidl-tools on the TI GitHub using a Bring Your Own Model flow.

Model artifacts are deployed through runtimes like ONNX Runtime, LiteRT (formerly tensorflow-lite) and TVM using TI Deep Learning (TIDL) as the backend software for hardware acceleration.

To deploy the model into an end-to-end vision application, start with edgeai-gst-apps, which composes the pipeline with multiple stages of hardware acceleration for pre-processing and post-processing the image, in addition to accelerating the AI model itself.

In addition to edgeai-gst-apps, we have built a demo for defect detection on small objects, ring terminals, using the YOLOX-Nano object detection network and edgeai-gst-apps as a baseline. The model was trained with Edge AI Studio to recognize good parts vs. multiple types of defects. Additional postprocessing and visualization shows statistics for the number of good and defective parts.