Programmable Solution- Implemented on a Lattice low power FPGA, the demo uses machine learning to identify different human faces. A CNN acceleration engine is trained to deliver accurate identification by extracting 256 16-bit characteristics from each registered face.
In-field Registration- The demo can register and identify faces without the need for retraining, removing the need for uploading images and lengthy retraining using a GPU.
Rapid Implementation- The demo’s development boards support RTL blocks for the 8-layer CNN accelerator, image sensor connectivity and setup, and image sensor processor and memory management or easy modification.
Features
- VGG8-like CNN trained to recognize human faces using measurement points
- In-field new face registration and identification without the need to retrain the network
- Supports performance of up to 30 frames per second
- Power consumption: 850 mW on Lattice ECP5 85K FPGAs and 200mW on Lattice CrossLink-NX 40K FPGAs
Support Discontinued - For Reference Only
