南京理工大学数字信号处理英文课件 Lecture20.ppt

274 * Lecture 20 Discrete Fourier Transform ( III ) Convolution via the Discrete Fourier Transform ? Principle ? Linear vs. circular convolution ? Applications with indefinite length signals A. Convolution via DFT: Principle According to the convolution : x[n] ? y[n] F X(e jω)Y( e jω) Then w[n] = x[n] ? y[n] may be computed as F－1 {F(x[n])F(y[n])} In practice: ? circular convolution ? Fourier transform implemented by DFT (fast algorithm) B. Linear vs. circular convolution When will w p[n] = w[n] ? Note: length of w[n] = length of x[n] (P ) +length of y[n] (L) －1 length of w p[n] = N Suppose: x[n] : 0≤n ≤P －1 ? 0≤m ≤P －1 y[n]: 0≤n ≤L －1 ? 0≤n － m ≤L －1 Rewrite the equations of linear and circular convolution ?If no time aliasing, w p[n] = w[n] B. Linear vs. circular convolution ( II ) ? x[n] y[n] N While w[n] = x[n] ? y[n] ? W (e jω) = X(e jω)Y (e jω), ?ω∈? For ω = 2π k /N, W (e2πj k/N ) = X[k]Y [k], k = 0, . . . , N － 1 with X[k] the N -point DFT of x[n]. Then W [k] = W (e 2πj k/N ) is the N -point DFT of both B. Linear vs. circular convolution ( II ) Left--- linear convolution; right---- circular convolution B. Linear vs. circular convolution ( III ) B. Linear vs. circular convolution ( IV ) B. Linear vs. circular convolution ( V ) Comment: ? when N≥L+P －1 , w p[n] = w[n]; ? when N ≤ L+P －1, w p[n] ≠ w[n]; the correct samples of linear and circular convolution: P+L － N－1≤ n ≤N －1 B. Linear vs. circular convolution ( V ) C. Block Convolution When L is indefinite, the signal is blocked in blocks of length L. ? Overlap-add method ? Overlap-save method ? non overlapping blocks of length L ? convolutions of blocks with other signal ? addition of contribution from different blocks C.1 Overlap-add * * *