Abstract: A novel image denoising method is proposed based on
multiscale wavelet thresholding (WT) and bilateral filtering (BF).
First, the image is decomposed into multiscale subbands by wavelet
transform. Then, from the top scale to the bottom scale, we
apply BF to the approximation subbands and WT to the detail
subbands. The filtered subbands are reconstructed back to approximation
subbands of the lower scale. Finally, subbands are
reconstructed in all the scales, and in this way the denoised image
is formed. Different from conventional methods such as WT and
BF, it can smooth the low-frequency noise efficiently. Experiment
results on the image Lena and Rice show that the peak signal-
to-noise ratio (PSNR) is improved by at least 3 dB and 0.7 dB
compared with using the WT and BF, respectively. In addition, the
computational time of the proposed method is almost comparable
with that of WT but much less than that of BF.
Key words: wavelet thresholding; bilateral filtering; multiscale;
image denoising

摘要：本文提出了一種基于多尺度小波閾值（WT）和雙邊濾波（BF）的圖像去噪方法。首先，圖像會利用小波變換被分解成多尺度子帶。然后，從頂部規(guī)模向底部規(guī)模，我們將BF應(yīng)用到近似子帶而將WT應(yīng)用到細(xì)節(jié)子帶。過濾后的子帶進行重構(gòu)后返回到了較低規(guī)模的近似子帶。最后，子帶在所有規(guī)模的范圍內(nèi)被重建后，并用這種方式對形成的圖像去噪。不同于使用WT和BF 的傳統(tǒng)方法，它可以有效地對低頻噪音進行去噪。從圖像上的實驗結(jié)果可以表明，莉娜 和 大米的信噪比（PSNR）對比使用了WT 和BF之后，分別提高了至少3分貝和0.7分貝。此外，該方法的計算時間幾乎可以與使用WT的方法相媲美但是遠小于使用BF的方法。