Hardware Implementation of Real-Time, High Performance, RCE-NN Based Face Recognition System

Hardware implementation of a real-time, highly accurate face recognition system (FRS) is proposed in this correspondence. Face images are acquired from a CMOS sensor camera connected to Field Programmable Gate Array (FPGA) based reconfigurable hardware board using Cam Link interface. We used contrast limited adaptive histogram equalization (CLAHE) for image contrast enhancement, discrete wavelet transform (DWT) to remove variable illumination & select appropriate subband and principal component analysis (PCA) with 35 principal components which is optimized for performance and speed. Finally, Restricted Coulomb Energy (RCE) based neural network (NN) classifier is used for face recognition. We have implemented the RCE based NN in FPGA and thus utilized the inherent parallelism effectively which is not possible with NN software implementation. The performance of our implementation is superior than face recognition software and hardware implementations, which are targeted to achieve higher recognition accuracy at faster rate using minimum computational resources. Our system recognizes a single image in real-time i.e. within 18 ms corresponding to 37 frames per second image capture. We have verified our proposed system with multiple standard face databases as well as using our own face data repository.