Abstract

The Kohonen Self-Organizing Map (K-SOM) has found applicability in a wide range of application areas. It has also proved to be successfully used in compressing digital images. The dominant point of the algorithm is that it is more suitable to realise in hardware platforms in order to speed up the overall operation. It, however, comes in exchange with the drawback of poorer resulting image quality compared to software implementation. This comes from the fact that both a codebook and a learning kernel within the hardware platform need to be approximated by an integer basis in order to limit the hardware utilization and speed up the execution time. In addition, the learning kernel which has a substantial impact on the quantized image quality is required to be a simple linear integer-based function. In this paper, we propose a hardware centric K-SOM quantizer algorithm which relies on a rational-based representation of the codebook and learning kernel. This extends the capability of the quantizer to accept an approximated non-linear learning kernel. The experimental results have proved that the quality of the outcome images is superior to predecessor implementations with an acceptable throughput and FPGA resource utilizations. The results show that the image quality, measured by the mean square error (MSE), has an average improvement ratio of 0.68 compared to the state-of-the-art integer representation K-SOM quantizer with respect to the standard test images while the proposed quantizer takes only 9% more of the FPGA resources. © 2011 Springer-Verlag.