"A Fast FPGA Implementation of a Unique Multi-level Tree-based Image Classifier"
Geoffrey Wall1, Faizal Iqbal1, Xiuwen Liu2, and Simon Foo1
1Florida A&M University - Florida State University
2Florida State University
Image recognition is a central problem in computer vision. Classification of images whether through statistical, neural network -based, or other kernel -based methods tend to require the calculation of transcendental functions such as the exponential. These functions can be approximated by using truncated Taylor series or by the use of CORDIC algorithms. These methods are often slow and take up many logic cells when implemented in an FPGA. We proposed and implemented a classification method for FPGAs which makes use of the large block ram and embedded multiply blocks available in many modern FPGAs. Our method uses a tree-based hierarchical approach in order to limit the amount of calculations required to reach a classification. Our classifier shows a significant speedup over traditional neural network -based classifiers with comparably accurate results on several large and diverse image datasets.
Keywords: Image recognition, tree-based classifier, neural networks, FPGAs
2005 MAPLD International Conference Home Page