In this paper, an efficient automatic diagnosis system for pneumonia classification is developed using extracted textural features obtained from appropriate wavelet transformation. For feature extraction and analysis in the classification of pneumonia, different wavelet families such as Db3.3, Rbio3.3, Rbio3.5, and Rbio 3.7 are explored. The optimum feature extraction for distinguishing pneumonia infected lungs from normal lungs comes from combining the Db3.3 and Rbio3.7 wavelet families. The features extracted from Db3.3 and Rbio3.7 wavelets are analyzed by feeding to different supervised learning classifiers. It is observed that SVM with RBF kernel is attaining maximum accuracy of 97.5% with sigma=2 in the classification. The RBF kernel, on the other hand, is hampered by its lengthy testing computation time. This paper introduces a novel backward elimination based SVM (BESVM) in order to reduce computation time. The suggested method's experimental findings demonstrate the trade-off between classification speed and performance. This was also noticed when targeting a real-time hardware software codesign FPGA environment. The amount of support vectors is optimized using the BESVM technique, resulting in a 30% reduction in resource utilization and a 590 ns delay while maintaining accuracy. In terms of area, latency, and hardware efficiency, the suggested BESVM-based hardware design demonstrates its efficacy.
展开▼
