An Android-based mobile AI Visual Inspection platform for industrial operations. It supports flexible multi-product configuration, combines cascaded inspection (coarse localization + precise recognition) with a step-based image-processing pipeline, and determines pass/fail from a single photo. Inspection data is uploaded to H3 BPM in real time for end-to-end digital management from part scanning through result traceability.
A two-stage architecture first localizes the ROI, then performs precise defect recognition. This avoids detail loss from high-resolution image compression and significantly improves small-object detection.
Supports 1D and 2D scanning with project-specific workflows (part number → serial number). A safety delay and short suppression window prevent repeated false scans.
Supports inspection modules for HCP4 gateway communication, 48V bracket clips, and 48V large/small pins. Centralized cloud configuration is pulled automatically by the app, enabling zero-code product changes on the shop floor.
Capture → EXIF normalization → crop → step-based rotation → inference. Supports 5 step types, including ROTATE_POST_CROP, with cloud-configured angles applied locally in real time.
Inspection results, on-site photos, and decision evidence are uploaded to H3 BPM in real time. Engineers can query records by serial number for end-to-end quality traceability from parts to finished products.
Admin, Engineer, and Operator roles provide 3 access levels, with binding for up to 3 devices. Mandatory cloud updates ensure the latest inspection logic is always used in production.
Industrial quality inspection has long depended on fixed cameras, introducing blind spots, high deployment costs, and limited flexibility. VisionQC brings cascaded Visual Inspection algorithms to mobile devices, allowing inspectors to sample multiple product types with one phone. It supports HCP4 gateway communication, 48V bracket clips, and 48V large/small pins, digitizing the workflow from part scanning through result traceability.
The first stage uses a lightweight YOLO model to locate the ROI in the full image. The second performs precise defect recognition on the cropped high-resolution image. This avoids detail loss from compressing the image to 640×640 and significantly improves small-object detection.
Scan part barcode → capture photo → normalize EXIF orientation → crop ROI → apply step-based rotation → run AI inference → determine result. Supports 5 step types (DETECT_COUNT, DETECT_CROP, ROI_CROP, ROTATE_POST_CROP, MANUAL_CONFIRM), with cloud configuration delivered in real time.
YOLOv8 models trained with PyTorch are exported to TFLite using an in-house conversion tool with FP16/FP32 support. NNAPI accelerates Android inference, while compatibility with [1,N,C] and [1,C,N] output formats makes models plug-and-play.
Inspection results, on-site photos, and decision evidence are uploaded to H3 BPM through row-level upserts keyed by ModuleCode+PointCode. Engineers can query historical records and images by serial number for end-to-end digital Traceability from part scanning through final disposition.