ImageMagick の例 --
画像比較

目次

ImageMagick の例 はじめに と 目次

画像比較の方法 -- 何が違うのか？

• compare プログラム
• 差分画像
• フリッカー比較
• アニメーションの比較

比較統計 -- どの程度違うのか？

部分画像の照合と位置特定 大きな画像の中から小さな画像を探す。

重複画像の検索 -- 同じ画像を2つ見つける

種類による画像のソート -- 比較のための画像分類

• グレースケール画像
• 線形カラー
• 純粋な白黒
• カラー画像
• 中間色のカラー
• テキスト vs 線画
• 実写 vs 漫画

特定の画像タイプの扱い

画像メトリクス -- 比較のための画像のフィンガープリント

ウェブカメラ -- 固定カメラで何が変更されたかを見つける

画像比較の方法

compare プログラム

magick compare bag_frame1.gif bag_frame2.gif magick compare.gif

  magick compare bag_frame1.gif bag_frame2.gif  x:
  magick compare bag_frame1.gif bag_frame2.gif  miff:- | display

magick compare bag_frame1.gif bag_frame2.gif \ -highlight-color SeaGreen magick compare_color.gif

magick compare bag_frame1.gif bag_frame2.gif \ -highlight-color SeaGreen -lowlight-color PaleGreen \ magick compare_colors.gif

magick compare bag_frame1.gif bag_frame2.gif \ -compose Src compare_src.gif

magick compare bag_frame1.gif bag_frame2.gif \ -compose Src -highlight-color White -lowlight-color Black \ magick compare_mask.gif

magick bag_frame1.gif bag_frame1.jpg magick compare bag_frame1.gif bag_frame1.jpg magick compare_lossy_jpeg.gif

magick compare -metric AE -fuzz 5% \ bag_frame1.gif bag_frame1.jpg magick compare_fuzz.gif

差分画像

magick composite bag_frame1.gif bag_frame1.jpg \ -compose difference difference_jpeg.gif

magick difference_jpeg.gif -auto-level difference_norm.gif

magick composite bag_frame1.gif bag_frame2.gif \ -compose difference difference_frames.gif

magick bag_frame1.gif bag_frame2.gif -compose difference -composite \ -threhold 0 -separate -evaluate-sequence Add \ difference_mask.gif

magick bag_frame1.gif bag_frame2.gif -compose difference -composite \ -colorspace Gray difference_gray.gif

magick bag_frame1.gif bag_frame2.gif -compose difference -composite \ -separate -evaluate-sequence add difference_add.gif

magick bag_frame1.gif bag_frame2.gif -compose difference -composite \ -evaluate Pow 2 -separate -evaluate-sequence Add -evaluate Pow 0.5 \ difference_vector.gif

magick bag_frame1.gif bag_frame2.gif -compose difference -composite \ -evaluate Pow 2 -evaluate divide 3 -separate \ -evaluate-sequence Add -evaluate Pow 0.5 \ difference_vector_scaled.gif

フリッカー比較

magick -delay 50 bag_frame1.gif bag_frame2.gif -loop 0 flicker_cmp.gif

アニメーションの比較

    magick \( anim1.gif -coalesce -append \) \
            \( anim2.gif -coalesce -append \) miff:- | \
      magick compare - miff:- |\
        magick - -crop 160x120 +repage anim_compare.gif

比較統計
2つの画像はどの程度異なるのでしょうか？

建設中

Statistics from difference image...

  The following outputs verbose information and extracts just the
  section containing the channel statistics of the image....

    magick image1 image2 -compose Difference -composite \
            -colorspace gray -verbose  info: |\
       sed -n '/statistics:/,/^  [^ ]/ p'

  The numbers in parenthesis (if present) are normalized values between
  zero and one, so that it is independent of the Q level of your IM.
  If you don't have these numbers, you should think of upgrading your IM.

  To get the average (mean) grey level as a percentage you can use this
  command...

     magick image1 image2 -compose Difference -composite \
           -colorspace gray -format '%[fx:mean*100]' info:

  For non-percentage you can use the even simplier..

     magick image1 image2 -compose Difference -composite \
           -colorspace gray -format '%[mean]' info:


Compare  Program  Statistics...

   You can get an actual average difference value using the -metric

     magick compare -metric MAE image1 image2 null: 2>&1

   Adding -verbose will provide more specific information about each separate
   channel.

      magick compare -verbose -metric MAE rose.jpg reconstruct.jpg null: 2>&1

      Image: rose.jpg
      Channel distortion: MAE
        red: 2282.91 (0.034835)
        green: 1853.99 (0.0282901)
        blue: 2008.67 (0.0306503)
        all: 1536.39 (0.0234439)

   Their are a number of different metrics to chose from.
   With the same set of test images (mostly the same)

   Number of pixels
      AE ...... Absolute Error count of the number of different pixels (0=equal)

                This value can be thresholded using a -fuzz setting to
                only count pixels that have a larger then the threshold.

                As of IM v6.4.3  the  -metric AE  count is -fuzz effected.
                so you can discount 'minor' differences from this count.

                magick -metric AE -fuzz 10% image1.png image2.png null:

                Which pixels are different can be seen using the output
                image (ignored in the above command).

                This is the ONLY metric which is 'fuzz' effected.

   Maximum Error (of any one pixel)
      PAE ..... Peak Absolute Error   (within a channel, for 3D color space)
      PSNR .... Peak Signal to noise ratio (used in image compression papers)
                The ratio of mean square difference to the maximum mean square
                that can exist between any two images, expressed as a decibel
                value.

                The higher the PSNR the closer the closer the images are, with
                a maximum difference occurring at 1.  A PSNR of 20 means
                differences are 1/100 of maximum.

   Average Error (over all pixels)
      MAE ..... Mean absolute error    (average channel error distance)
      MSE ..... Mean squared error     (averaged squared error distance)
      RMSE .... (sq)root mean squared error -- IE:  sqrt(MSE)


   Specialized metrics
      MEPP .... Normalized Mean Error AND Normalized Maximum Error
                These should directly related to the '-fuzz' factor,
                for images without transparency.

                With transparency, makes this difficult the mask should
                effect the number of pixels compared, and thus the 'mean'
                but this is currently not done.

      FUZZ      fuzz factor difference taking transparency into account

      NCC       normalized cross correlation (1 = similar)

   I produced the following results on my test images...

    _metric_|__low_Q_jpeg__|__black_vs_white__
     PSNR   | 29.6504      | 0
     PAE    | 63479        | 65535
     MAE    | 137.478      | 65535
     MSE    | 4.65489e+06  | 4.29484e+09
     RMSE   | 2157.52      | 65535


   The first column of numbers is a compare of images with low-quality JPEG
   differences, where the test image was read in and saved with a very low
   -quality setting.

   The second "black vs white", is a compare of a solid black image verses
   a solid white image.  If the 'average color' of the image is ignored
   by the comparision then the resulting value will be very small.  This
   seems only to be the case with the PSNR metric, as all others produced
   a maximum difference value.

   The e+06 is scientific notation, on how many places to shift the
   decimal point.  EG:   4.65489e+06  -->  4,654,890.0
   Thus is equal to about 4 million, and is the square of 2157.52

   WARNING: numbers are dependant on the IM Quality (Q) levels set at compile
   time. The higher the quality the larger the numbers. Only PSNR should be
   unaffected by this.  For this reason IM also gives you a 'normalized'
   result that is uneffected by the compile time quality setting, though may
   still have minor 'quantum' or 'interger rounding' effects.

   I have NOT figured out if there are any of the existing "-define" options
   usable the "compare" function.


   NOTE for opaque colors AE -fuzz  and RMSE distances are equivelent.
   HOWEVER,  when transparent colors are involved AE fuzz factor testing
   will treat two different fully-transparent colors as being the same
   while RMSE will treate them as being different!

   For example...
   To AE fully-transparent white and fully-transparent black are the same.

     magick compare -metric AE xc:#0000 xc:#FFF0 null:
     0

   But to RMSE they are VERY different

     magick compare -metric RMSE xc:#0000 xc:#FFF0 null:
     56755 (0.866025)

Dissimilarity-threshold

  If you get a 'too different' error,  you can disable that using...
      -dissimilarity-threshold 1.0

  But what is this threshold?

部分画像と形状の照合

建設中

Using "compare -subimage-search" option...

  magick compare -subimage-search  large_image.png  sub-image.png  results-%d.png

  This produces two images
    results-0.png
        which displays the matching location
    results-1.png
        which is a map of possible top-left corner locations showing how well
        the sub-image matches at that location.

  Note the second image is smaller, as it is only top-left corner locations.
  As such its size is   large_image - small_image + 1

  The search however is based on a difference of color vectors, so produces
  a very accurate color comparison.

  The search basically does a compare of the small image at EVERY possible
  location in the larger image.  As such it is is slow! very very slow..

  The best idea is to compare a very very SMALL sub-image to find possible
  locations, than use that to then do a difference compare at each possible
  location for a more accurate match.

  Have a look at the script
    https://imagemagick.dokyumento.jp/Usage/scripts/overlap
  and associated discussion
    Overlapped Images
  Which looks at locating 'high entropy' sub-images of one image to search
  for posible matches in a second image so the overlap offset between the
  two images can be discovered, and the images merged into a larger image.

  Another discussion uses sub-image searches to find tiling patterns in
  larger images, with the goal of generating tilable images
    Stitching image over a canvas


  Example using RMSE and the new -grayscale function to merge the
  separate color difference channel results into a final image

    magick large_image.png small_image.png miff:- |
      magick compare -metric RMSE -subimage-search - miff:- |
        magick - -delete 0 -grayscale MS show:


Similarity Threshold

  As many time people are only interested in the first match that matches.
  As soon at this 'good' match is found, there is no need to continue
  searching for another match.  The -similarity-metric defines what you
  would regard as a good match.

  A "-similarity-threshold 0.0" will abort on the very first 'perfect' match
  found, while "-similarity-threshold 1.0"  (the default) will never match and
  will search every posible point.  A value between will set a color only
  'fuzz' factor on what you would find an acceptable match.

  Note that if the sub-image search is aborted, the second 'map' image will
  only contain a partial result, only showing the results up until the comapre
  aborted its search).


Some Basic Sub-Image Search Examples....

  Grab a screen shot of a terminal window ("screen.png"),
  and crop out an image of a single letter or word ("letter.png").

  Just report first match.... for speed,
  immeditally abort after finding that first match.
  Don't bother outputing the incomplete image results.

     magick compare -subimage-search -metric AE -similarity-threshold 1.0 \
                   screen.png letter.png null: 2>&1

  NOTE speed will be highly dependant on where in the image that first
  match is found.

  Find all occurances of exactly that image,
  as an image (white dots on matches, black elsewhere)

     magick compare -subimage-search -metric AE \
                   screen.png letter.png miff:- 2>/dev/null |
       magick - -delete 0 show:

  Extract a list of the coordinates of all matching letters (white dots)
  (as an enumerated pixel list, ignoring anything black)

     magick compare -subimage-search -metric AE \
                   screen.png letter.png miff:-  2>/dev/null |
       magick - -delete 0 txt:- | grep -v '#000000'

  Just the coordinate list

     magick compare -subimage-search -metric AE \
                   screen.png letter.png miff:-  2>/dev/null |
       magick - -delete 0 txt:- | sed -n '/#FFFFFF/s/:.*//p'



NON-ImageMagick sub-image search solutions...

  "visgrep" from the "xautomation" package.

    This is much simpler sub-image search program, that only outputs a
    list of coordinates for the matches (or even multiple sub-image matches).
    Because it is so much simpler (for near exact matching) and not trying
    to generate 'result images' for further study, it is also a LOT FASTER.

    For example...
  
      visgrep screen.png letter.png

    Timed results
      using "compare" to get just the first match        0.21 seconds
      using "compare" to get a 'results image'           1.56 seconds
        ditto, but extracting the coordinate list        1.76 seconds
      using "visgrep" to get all matching coordinates    0.09 seconds



Other Methods of sub-image searching....

HitAndMiss Morphology

  This is essentually a binary match, where you define what pixels much be
  'background' and what must be forground.  However it also allows you to
  define areas where you don't care if the result is a foregorund or
  background.

  Basically a binary pattern search method.

Correlate (a Convolve variant)

  This is similar to Hit and Miss but using greyscale values.  Positive values
  for forground and negative values for background, and zero for don't care.
  It is however limited to grayscale images.

  See Correlation and Shape Searching.

  Both of these are basically just as slow as the previous sub-image compare,
  but less accurate with regards to colors.  However it's ability to specify
  specify a shape (don't care areas) to the sub-image makes then useful as
  a search method.

  However you need to magick the sub-image into a 'kernel', or array of
  floating point values, rather than as an actual image.


FFT Convolve (NCC)

  Fast Fourier Transforms is a slow operator, but usually many orders of
  magnitude faster than the previous two methods use.  The reason is that
  a convolution in the frequency domain is just a direct pixel by pixel
  multiplication.

  The 'Convolve' method, can be converted into a 'Correlate', simply by
  rotating the sub-image being searched for by 180 degrees.
  See Correlate.

  Basically by converting images into the 'frequency' domain, you can do
  a sub-image search, very very quickly, compared to the previous, especially
  with larger sub-images that can be the same size as the original image!

  This I believe has been added as a NCC compare metric.



Peak Finding and extracting (for near partial matches)...

  Once you have compared the image you will typically have a 'probably map'
  of some kind which defines how 'perfect' the match was.

  What you want to do now is to find the best match, or perhaps multiple
  matches in the image.  That is, you want to locate the major 'peaks'
  in the resulting map, and extract actual locations.

  * Using a Laplacian Convolution Kernel

    To get results you need to find the 'peaks' in the image, not
    necessarily the brightest points either. You can get this by convolving
    the image so as to subtract the average of the surrounding pixels from
    the central pixel.  As we only want positive results, a bias removes the
    negative results.

      magick mandril3_ncc1.png \
              -bias -100% -convolve Laplacian:0 result.png

    Thresholding and using it as a mask, and we can extract just those pixels.

      magick mandril3_ncc1.png \
              \( +clone -bias -100% -convolve Laplacian:0 -threshold 50% \) \
              -compose multiply -composite \
              txt:- | grep -v black

    The problem is you can get a cluster of points at a peak, rather than
    a definitive pixel, especially for two peak pixel surrounded by very low
    values.

  * Using a Peaks Hit and Miss Morphology Kernel

      magick mandril3_ncc1.png \
              -morphology HMT Peaks:1.5 result.png

    The problem is that this may produce no result if you get two peak pixels
    with exactly the same value (no gap between foreground and background)

    However there are other 'peak' kernels that will still locate such a peak
    cluster.

  * Dilate and compare

    Dilate (expand maximum values) the image 3 times then compare it to the
    original image.  Any peak within the area of dilated kernel size (7 pixel
    square) will remain the same value. Set all pixels that show a
    difference to pixels to zero.

    Method by HugoRune  (IM discussion topic 14491)

  * Looped match and remove.

    Basically find the highest pixel value, note it. Then mask all pixels in
    an area around that peak, and repeat until some limit (number points or
    threshold) is reached.

    See a shell script implementation of this in  Fred Weinhaus's script
    "maxima"

    This does not look at finding the center of large 'cluster' of near equal
    valued pixels, though this would be very rare in real images.

  * Sub-pixel locating

    If the peak is not an exact pixel, but could conceivably be a sub-pixel
    location (between pixels) then some form of pattern match (gaussian curve
    fit) in the area of the peak may let you locate the peak to a sub-pixel
    coordinate.

    This may be more important in image registration for parorama stitching,
    especially when you are not using a large number points to get a best-fit
    average of the perspective overlay.

  * Finding a tile pattern in an image

    When you have all the points, a search for a repeating pattern (similar
    vector distances between multiple peaks) should point out some form of
    tiling structure.


Improving the Sub-Image Matching...

  The major problem with Correlate, (or the fast  FFT correlate, which is the
  same thing) is that it has absolutely no understanding of color.

  Correlation (or convolve) is purely a mathematical technique that is used
  against a set of values.  With images that means it is only applied
  against the individual channels of an image, and NOT with vector color
  distances.


  While compare actually does real comparing of color vectors.  This will find
  shapes better than correlate but is much much slower.

  As such to make proper use of correlate you should magick your images
  (before hand for speed, or afterward against results) to try and highlight
  the color differences in the image as a greyscale 'correaltion' image.

  ASIDE: Use -channel to limit operations to one greyscale channel will
  improve speed.  In IMv7 greyscaling will reduce images to one channel so
  will gain speed improvements automatically.

  For example instead of intensity, you may get a better foreground
  / background differentiation, by extracting the  Hue of an image.
  Though you may need to color rotate the hue's if there is a lot of red
  in the sub-image being searched for.

  See the examples of HSL and HSB, channel separation, to see this problem.
    https://imagemagick.dokyumento.jp/Usage/color_basics/#separate

  Another greyscaling method that should work very well is to do edge
  detection on the two images.  This will highlight the boundaries and shape,
  which is typically much more important than any smooth gradient or color
  changes in the image.

  For examples of Edge detection methods see
    https://imagemagick.dokyumento.jp/Usage/convolve/#edgedet

  You may like to also look at directional or compass type edge detection.

  Basically Anything that will enhance the shape for your specific case is
  a good idea.  Just apply it to BOTH images before correlating them.


Scale and Rotation Invariant Matching...

  * position independence...
  * matching rotated sub-image (angle independent)
  * matching resized sub-images  (size independent)
  * Both size and angle independence


--------------

Other more specific image matching..

Matching Lines...

  Hough Algorithm

Matching Circles...

  Hough Algorithm Variant

Matching Faces

  A combination of the above.

重複画像の検索

同一ファイル

  md5sum * | sort | awk {'print $2 " " $1'}  | uniq -Df 1

IM画像署名

  magick identify -quiet -format "%#" images...

直接比較

画像分類

サムネイル比較

画像メトリクス

類似度または距離

• 直接しきい値、または最大差（チェビシェフ距離）
  1つのメトリクスで最大の差がある画像を比較します。
  しきい値は、多次元メトリクス空間で類似した画像の超立方体を生成します。もちろん、画像の違いは1つのメトリクスのみに基づいており、すべてのメトリクスには基づいていません。
• 平均差（平均距離、平均マンハッタン距離）
  すべての差を合計し、オプションでメトリクスの数で除算します。
  これは、都市グリッドを移動するために必要な距離に相当するため、2つのメトリクス間のマンハッタン距離としても知られています。すべてのメトリクスが等しく寄与するため、予想よりも「近い」ものに見えます。空間では、このメトリクスのしきい値はダイヤモンドのような形状になります。
• ユークリッド（ピタゴラス）距離
  または、メトリクス空間におけるメトリクス間の直接的なベクトルの距離。
  より多くのメトリクスが関与すると、値は大きくなる傾向があります。ただし、1つのメトリクスが大きな差を生み出すと、他のメトリクスよりも寄与が大きくなる傾向があります。しきい値は、メトリクス空間に球形のボリュームを生成します。
• 数学的誤差/データフィットまたは（慣性モーメント???）
  すべての差の二乗の合計を求め、平方根を取得します。
  これは、数学的な曲線が特定のデータセットにどれだけ近いかを計算するためによく使用されますが、画像メトリクスを比較するためにも使用できます。
  これは、最適な非ベクトル距離測定を提供するようです。
• ベクトル角度
  画像のメトリクスによって作成されたベクトル空間の中心からの2つの線の間の角度を見つけます。これにより、2つの画像に適用された可能性があるコントラストや画像強調の効果が除去されるはずです。
  まだテストされていません
• ベクトル距離
  メトリクスのすべての個々のカラーベクトルが同じ方向にある線画またはグレースケール画像の場合、画像の平均色からのメトリクスの相対距離がおそらくより重要です。最大距離に対する距離を正規化すると、コントラストの影響が軽減される可能性があります。
  つまり、これは線画画像の比較方法です。
  まだテストされていません
• クラスタ分析
  すべてのメトリクスがプロットされ、多次元空間内の類似したクラスターにグループ化されます。優れたクラスタリングパッケージでは、クラスタリングを生成しないメトリクスを発見して除外することもできる場合があります。
  まだテストされていません

人間の検証

• 画像を並べて表示する
• 元のサイズで、オプションで共通の「スケーリングされた」サイズで、2つの画像を非常に高速に切り替える。
• 画像を一致させようとするために、異なるスケールで変換された画像を切り替えたり、重ねたりする。
• 一致する画像と同じディレクトリ（ソース）またはおそらく同じクラスター（他の近接一致）内の他の画像を見て、各画像を個別に処理するのではなく、グループ全体を処理する。
• 画像を並べ替えたり、他の画像を拒否したりするために、2つ（またはそれ以上）のディレクトリ間で、画像のファイル名の変更、移動、置換、削除、コピーを行う。
• など...

クロスタイプ画像比較

重複画像を見つけるためのまとめ

   Generate/Cache  Image Types and Metrics
     -> Compare metrics and cluster images.
       -> compare images in cluster for matches
         -> group into sets of matching images (by source directory)
           -> human verification

タイプ別の画像ソート

• 白黒の線画またはテキスト画像（ほぼ全てが単色）
• 2つの基本色で構成される画像 - 同じくらい（パターン画像？）
• グレースケールのアーティストの描画（多くの陰影がある）
• 線形カラー画像（色はグラデーションを形成しますが、白黒からはではありません）
• 大きな単色領域のある漫画のようなカラー画像
• 陰影のある色の領域のある実写画像
• 画像には、注釈付きのテキストまたはロゴオーバーレイが含まれています。（単色のスパイク）

• 画像全体の平均色
• 画像内の主要な色
• 画像の前景/背景色。

グレースケール画像

magick rose: granite: -colorspace HCL \ -format '%M avg=%[fx:mean.g] peak=%[fx:maxima.g]\n' info:

画像は線形カラーですか

純粋な白黒画像

magick wmark_dragon.jpg -solarize 50% -colorspace Gray wmark_bw_test.png magick identify -verbose -alpha off wmark_bw_test.png | \ sed -n '/Histogram/q; /Colormap/q; /statistics:/,$ p' > wmark_stats.txt

magick granite: granite.jpg magick granite.jpg -solarize 50% -colorspace Gray granite_bw_test.png magick identify -verbose -alpha off granite_bw_test.png | \ sed -n '/Histogram/q; /Colormap/q; /statistics:/,$ p' > granite_stats.txt

スポットカラー画像

• 彩度または彩度チャネルを持つカラースペースに画像を変換する
• 画像を1:50（2％）の比率で小さくサイズ変更する（例：色の「スポットサイズ」）
• 最大彩度/彩度値で閾値処理する

中間調カラー画像

テキスト vs 線画

実写風 vs 漫画風

• すべてのピクセルの色を書き出す
• 色でソートする
• 色ごとのピクセル数を書き出す
• ピクセル数でソートする
• 画像内のピクセルの半分を占めるまで、リストを順に処理する。
• #ピクセル >>> #色 の場合、漫画風です。

特定の画像タイプの処理

不良なスキャンまたは印刷物

FUTURE -- image example needed for testing
    magick bad_printout.png -crop 0x1+0+0 -evaluate-sequence mean \
            -threshold 50% -format %c histogram:info:-

faster method but needs image height (assumed to be 1024)
    magick bad_printout.png -scale 1024x1 \
            -threshold 50% -format %c histogram:info:-

空白のファックス

FUTURE -- image example needed for testing
    magick blank_fax.png -threshold 50% -format %c histogram:info:-

FUTURE -- image example needed for testing

スパム画像

電子メールスパム画像

画像メトリクス、比較する画像をすばやく見つける

画像の平均色

You can use -scale to get an average color of an image, however I also suggest
you remove the outside borders of the image  to reduce the effect of
any 'fluff' that may have been added around the image.

    magick image.png  -gravity center -crop 70x70%+0+0 \
            -scale 1x1\! -depth 8 txt:-

Alternatively to get 'weighted centroid' color, based on color clustering,
rather than an average, you can use -colors

    magick rose: -colors 1 -crop 1x1+0+0 -depth 8 -format '%[pixel:s]' info:-
    rgb(146,89,80)

画像の主な色

 FUTURE example 

境界色

ヒストグラム - 一般的な色のマッチング

前景/背景色の分離

  magick rose: -median 5 +dither -colors 2 \
          -depth 8 -format %c histogram:info:- 

  magick rose: -median 5 +dither -colors 2 \
          -trim +repage  -gravity center -crop 50% \
          -depth 8 -format %c histogram:info:- 

平均カラーマトリックス

  magick logo: -scale 3x3\! -compress none -depth 8 ppm:- |\
    sed '/^#/d' | tail -n +4

  251 241 240 245 234 231 229 233 236 254 254 254
  192 196 204 231 231 231 255 255 255 211 221 231
  188 196 210

  magick logo: -gravity center -crop 80% -scale 3x3\! \
          -compress none -depth 16 ppm:- |   sed '/^#/d' | tail -n +4

  63999 59442 58776 62326 58785 58178 51740 54203 54965 65277 65262 65166
  45674 47023 49782 56375 55648 55601 65535 65535 65535 52406 55842 58941
  44635 48423 52881

色差マトリックス

  magick logo: -gravity center -crop 80% -scale 3x3\! -fx '.5+u-p{1,1}' \
          -compress none -depth 16 ppm:- | sed '/^#/d' | tail -n +4

  51093 45187 41761 49419 44529 41163 38834 39947 37950 52371 51007 48152
  32767 32767 32767 43469 41393 38587 52629 51279 48521 39500 41587 41926
  31729 34168 35867

  magick logo: -scale 3x3\! \( +clone -scale 1x1 \) -fx '.5+u-v.p{0,0}' \
          -compress none ppm:- | sed '/^#/d' | tail -n +4

  38604 35917 34642 37011 33949 32441 32839 33841 33649 39447 39259 38369
  23358 24377 25436 33538 33174 32426 39612 39434 38605 28225 30576 32319
  22271 24381 27021

近傍の差

知覚ハッシュ

より良い画像のマッチング

セグメンテーションカラー

  magick logo: -median 10 -segment 1x1 \
          +dither -scale 100x100\! segment_image.gif

  magick logo: -scale 100x100\! -median 3 \
          -quantize YIQ +dither -colors 3 segment_image.gif

無色のエッジ比較

  magick logo: -scale 100x100\! -median 3 \
          -quantize YIQ +dither -colors 3 -edge 1 \
          -colorspace gray -blur 0x1   outline_image.gif

  magick logo: -scale 100x100\! -median 3 \
          -quantize YIQ +dither -colors 3 \
          -lat 3x3-5% -negate \
          -colorspace gray -blur 0x1  outline_image.gif

Webカメラ
固定カメラで何が変わったか

建設中