PCB Manufacturer Cuts False Rejects by 35% with Edge AI Vision System That Filters AOI Output in Real Time

Before building anything, we needed to understand exactly why the existing AOI was over-rejecting. Our team spent two weeks on-site, collecting thousands of images that the AOI had flagged as defective.

We worked with the client’s QC team to manually classify each image: real defect or false positive. The patterns became clear – lighting variations and lead-free solder reflections accounted for the majority of false flags. This dataset became the foundation for model training.

The key insight from our analysis was that lighting normalization had to happen before inference. Software-only solutions couldn’t keep up with line speed while also correcting for ambient lighting changes.

The NXP i.MX 8M Plus solved this with hardware ISPs that handle lens distortion correction and white balance at full frame rate. The dedicated NPU then runs inference on the already-clean images. This split architecture – hardware preprocessing plus neural acceleration – was essential for real-time performance.

Generic object detection models couldn’t reliably distinguish between actual solder defects and the complex reflections that lead-free joints produce. We trained a MobileNet SSD architecture specifically on the client’s defect types: solder bridges, tombstoning, insufficient wetting, and cold joints.

Using NXP’s eIQ Toolkit, we optimized the model for NPU inference, achieving classification speeds that matched the conveyor rate. The model learned to recognize the specific reflection patterns of the client’s lead-free solder formulation – the exact patterns that had been fooling the standard AOI.

The final step was connecting the edge system to the existing conveyor controls. We implemented direct GPIO signaling to the line’s PLC, enabling automatic board sorting without operator intervention.

The workflow is simple: when the primary AOI flags a board, our system re-analyzes the image. If the ML model classifies it as a false positive, it signals the PLC to continue the board down the pass line. Only boards confirmed as truly defective are diverted for manual review – eliminating over 35% of the false rejects that had been clogging the review queue.