Vision AI at the Edge for Industrial Inspection
summary
Neruala vision AI runs on Qualcomm QCS6490/QCM6490 modules with Windows 11 IoT Enterprise to detect conveyor‑belt defects in about 10 classes; model training uses standard datasets, compression with the AI Model Efficiency Toolkit, and compilation/INT8 quantization via the Qualcomm Neural Processing SDK. Quantized models keep accuracy within a few percent of FP32, deliver sub‑millisecond inference latency on the reference board, and handle real samples like wood slabs and carpets while using the chip’s triple ISP and integrated 5G.
AI-written summary · about
chapters (11)
- 0:00Vision AI at the edge
- 1:38Industrial inspection apps
- 3:08Qualcomm’s edge approach
- 6:43On-device and edge AI box
- 8:54QCS/QCM 6490 chipset
- 12:11Unified AI software stack
- 15:25Industrial defect detection
- 17:34Model dev and deployment
- 20:44Anomaly detection results
- 22:53Accuracy retention proof
- 23:35Developer resources
related talks
08mUnlocking AI at Scale for PyTorch Developers
54mHarnessing On-Device AI to Transform the Smartphone Experience
50mAI acceleration at the edge
16mRVEdge-Vision: A Fully Open,Ultra-Efficient On-Device AI Platform for Smart Eyewear
18mHow to turn your ideas in AI into products
31mRealtime Edge AI SoC for Computing Vision Applications
14mEmbedded Executive LIVE
42mThe Stakes of Green AI for Embedded Systems