FPGA implementation of a frequency adaptive learning SOFM for digital color still imaging

This paper presents an efficient architecture of a Kohonen self-organizing feature map (SOFM) based on a new frequency adaptive learning algorithm. For scalability, a broadcast architecture is adopted with homogenous synapses composed of shift register, counter, accumulator and a special MIN_FIND unit. The MIN_FIND unit speeds up the search for neurons with minimal attributes. Dead neurons are reinitialized at preset intervals to improve their adaptation. The proposed SOFM architecture is prototyped on a Xilinx Virtex FPGA. Experimental results show that a 64-neuron network uses 99% of a 1000 Kgate FPGA and the maximum frequency of operation is 25.34 MHz. A 512×512 pixel color image can be quantized in about 1.38 s at 25 MHz clock rate without the use of subsampling.