Algorithmic transformations in the implementation of K- means clustering on reconfigurable hardware

摘要: In mapping the k-means algorithm to FPGA hardware, we examined level transforms that dramatically increased achievable parallelism. We apply multi-spectral and hyper-spectral images, which have tens hundreds of channels per pixel data. K-means is an iterative assigns each a label indicating K clusters belongs to.K-means common solution segmentation multi-dimensional The standard software implementation uses floating-point arithmetic Euclidean distances. Floating point multiplication-heavy distance calculation are fine on general purpose processor, but they large area speed penalties when implemented FPGA. order get best performance FPGA, needs be transformed eliminate these operations. effects using two other measures, Manhattan Max, do not require multipliers. also fixed precision truncated bit widths in algorithm.It important explore algorithmic tradeoffs reconfigurable hardware. A direct translation would result very inefficient use hardware resources. Analysis data necessary for more efficient implementation. Our resulting exhibits approximately 200 times up over