Baxermux的攝影作品Blog

之前有介紹到一款極為強大.效果極端優異的去雜訊軟體Noiseware( http://www.wretch.cc/blog/baxermux/349909 ) ,但是去除雜訊後品質到底提升多少? 現在利用之前所開發的PSNR比較軟體 http://www.sendspace.com/file/gj8jim ,來進行數據客觀化的比較.

關於這個測試需要注意到幾個問題點

1.一定要很注意拍攝每張範例檔的時候對焦位置與拍攝距離視角有沒有任何輕微的改變 (其實即使把相機放在桌上或是固定在腳架上,稍微一大力按下設定鈕就可能造成相機移位了... )

2.亮度程度的偏移也是一個問題,好比說測光下需要1/120秒曝光,iso提高一倍後,理論上需要1/240秒,但是因為相機可能沒有1/240秒這個快門數值,找最接近的1/250秒下去曝光,就有可能造成影響亮度的誤差偏移,不過這個影響性算是甚微,主要造成PSNR下降主因還是雜訊.

接著公佈幾個實驗的測試數據 (使用GF1 ISO 100拍攝範例檔當成標準範例 )

原始拍攝raw檔下載 http://www.sendspace.com/file/a4o4me  , 使用LightRoom3輸出16bits tif檔,後續處理(轉成bmp檔和縮圖)使用光影魔術手,bmp轉成程式可讀格式使用UleadPhotoimpact 8轉換.

實驗一. 比較不同ISO跟ISO 100的PSNR數據 (從上到下分別是ISO 200,400,800,1600,3200)

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120026.bmp
==========
PSNR of P1120025.bmp & P1120026.bmp

PSNR R:32.8047
PSNR G:33.1078
PSNR B:34.0138
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120027.bmp
==========
PSNR of P1120025.bmp & P1120027.bmp

PSNR R:32.6542
PSNR G:32.5736
PSNR B:32.0713
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120028.bmp
==========
PSNR of P1120025.bmp & P1120028.bmp

PSNR R:30.4396
PSNR G:30.8031
PSNR B:30.057
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120029.bmp
==========
PSNR of P1120025.bmp & P1120029.bmp

PSNR R:27.2699
PSNR G:27.7895
PSNR B:26.5547
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120030.bmp
==========
PSNR of P1120025.bmp & P1120030.bmp

PSNR R:23.9832
PSNR G:24.3792
PSNR B:22.8062
==========

現在共同的定義是psnr在30db以上都還算人眼尚可接受的高畫質範圍 ,低於30 db則是人眼不可忍受的畫質,ISO 800幾乎在最底線位置,上次測試則是低於30db,畢竟每次拍攝的範例和環境情況.誤差因素等等都會造成數據的偏移.可以看到iso 1600跟3200是糟糕到一個誇張的程度.

實驗二. 將檔案經過NoiseWare處理去雜訊過比較

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120026_filtered.bmp
==========
PSNR of P1120025.bmp & P1120026_filtered.bmp

PSNR R:32.805
PSNR G:32.963
PSNR B:34.1568
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120027_filtered.bmp
==========
PSNR of P1120025.bmp & P1120027_filtered.bmp

PSNR R:33.5057
PSNR G:33.2025
PSNR B:32.9106
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120028_filtered.bmp
==========
PSNR of P1120025.bmp & P1120028_filtered.bmp

PSNR R:32.5234
PSNR G:32.7811
PSNR B:32.2576
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120029_filtered.bmp
==========
PSNR of P1120025.bmp & P1120029_filtered.bmp

PSNR R:30.8842
PSNR G:31.8213
PSNR B:29.9382
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025.bmp P1120030_filtered.bmp
==========
PSNR of P1120025.bmp & P1120030_filtered.bmp

PSNR R:28.5125
PSNR G:29.5612
PSNR B:26.258
==========

經過去噪軟體處理除了在極低iso下有適得其反效過外(但是也沒說糟糕很多,只是db值有輕微下降,在iso 200情況下),其他iso 400,800,1600,3200都有進步,尤其在極高iso下進度幅度較高,iso 1600除了藍色頻道的psnr還是低於30db外,其他都達到堪用水準,至於iso 1600則是改善許多,但是還是達不到堪用的畫質.

實驗三. 原圖縮成800x600之下比較

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120026-s.bmp
==========
PSNR of P1120025-s.bmp & P1120026-s.bmp

PSNR R:35.246
PSNR G:35.5587
PSNR B:37.333
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120027-s.bmp
==========
PSNR of P1120025-s.bmp & P1120027-s.bmp

PSNR R:37.1806
PSNR G:36.4818
PSNR B:35.4006
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120028-s.bmp
==========
PSNR of P1120025-s.bmp & P1120028-s.bmp

PSNR R:37.1605
PSNR G:37.9044
PSNR B:35.6227
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120029-s.bmp
==========
PSNR of P1120025-s.bmp & P1120029-s.bmp

PSNR R:34.2055
PSNR G:36.3132
PSNR B:31.7995
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120030-s.bmp
==========
PSNR of P1120025-s.bmp & P1120030-s.bmp

PSNR R:30.8078
PSNR G:32.5593
PSNR B:27.3106
==========

沒錯...所謂縮圖之下無爛圖,就是這個道理,縮圖本身也會捨去雜訊.....

實驗四. 去雜訊的圖片再進行縮圖比較

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120026_filtered-s.bmp
==========
PSNR of P1120025-s.bmp & P1120026_filtered-s.bmp

PSNR R:35.0782
PSNR G:35.3017
PSNR B:37.4167
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120027_filtered-s.bmp
==========
PSNR of P1120025-s.bmp & P1120027_filtered-s.bmp

PSNR R:36.9822
PSNR G:36.2346
PSNR B:35.5563
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120028_filtered-s.bmp
==========
PSNR of P1120025-s.bmp & P1120028_filtered-s.bmp

PSNR R:37.2344
PSNR G:37.802
PSNR B:36.1517
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120029_filtered-s.bmp
==========
PSNR of P1120025-s.bmp & P1120029_filtered-s.bmp

PSNR R:34.5435
PSNR G:36.778
PSNR B:32.3708
==========

F:\gf1\20110129-iso-test\bmp>psnr.exe P1120025-s.bmp P1120030_filtered-s.bmp
==========
PSNR of P1120025-s.bmp & P1120030_filtered-s.bmp

PSNR R:31.506
PSNR G:33.6301
PSNR B:27.8126
==========

也是有進步...

那麼結論是?以GF1來說ISO 1600的確非常不甚理想(其實ISO 800就幾乎快要達到不可忍受的底線),不過經過去噪軟體處理,再加上縮圖後,其實已經算是達到堪用品質,在光圈.快門.ISO的選擇調配下,再考慮到去噪與縮圖後的效果程度,該怎麼分配這些硬體設定條件,我想會有更多選擇.