This reference design implements Convolutional Neural Network(CNN) based human face identification application on a low power Lattice FPGA using an image sensor. The training process is completed on a GPU-powered machine to sharpen the CNN to detect points of reference on a human face and measure them to distinguish the differences between people. This design can be used for identification of other objects by modifying the training database.
The hardware-based reference design includes an SPI, DDR memory interface blocks, an image signal processing engine, eight CNN acceleration engines, and a counting and overlay engine to show registration and identification results.
When the design is deployed on the FPGA, a person can register their face during the registration phase, using 256 different 16-bit values representing distinguishing facial characterizes are extracted and stored. In the identification phase, a registered person’ face can be identified using the hardware 256 16-bit values are extracted and compared to the stored list of values for verification.
Features
- VGG8 like – 8x (Convolution, Batch Normalization) + 4x Pooling + 1 fully connected CNN
- Network is trained with around 1 million images including augmentation for various conditions
- Performance of up to 30 frames per second
- Power consumption: 850 mW on ECP5 85K s and 200mW on CrossLink-NX 40K
