부산대학교 도서관

This paper proposes selective compressed sensing, an extension of the conventional compressed sensing that effectively compresses the frequency band(s) of interest while maintaining a backward compatibility with the compressed sensing. We exploit the fact that several sensing applications are interested in analyzing specific frequency bands only. In contrast to most conventional researches that use a random matrix as a measurement matrix, we present an optimized measurement matrix using the Walsh-Hadamard matrix and demonstrate that our selective compressed sensing method can reduce the amount of sensing data required to be transmitted. Compressed sensing can reconstruct all frequencies evenly from the raw data; the accuracy of the reconstructed data is primarily determined by the sparsity of the measurement signal. However, the compression process in the proposed selective compressed sensing method bypasses the calculation of unnecessary Fourier frequency band signals during the measurement. To reduce computational instructions and enable real-time compression by the resource-constrained microprocessors, we introduce two new techniques: the measurement matrix separations and the fast Walsh-Hadamard transform (FWHT) sorted by sequency. We demonstrate that this compression algorithm can reduce the number of instructions to approximately 1/512 of that of the compressed sensing to extract 40% of the whole spectrum from the raw data, when the data chunk size is 4096 bytes.