Imscript is a collection of simple programs for doing image processing in a Unix environment.
This is just vanilla unix philosophy with some specificities for images:
Notice that imscript programs will read any common image format, but will only write asc, pnm, png, and tiff.
1. Filters: read one image and write one image of the same size - blur: convolve an image by a shift-invariant user-specified kernel - morsi: apply a morphological operation with a user-specified element - qeasy: re-scale the dynamic range of an image - qauto: re-scale the dynamic range of an image, automatically - palette: colorize a grayscale image using a palette - dither: binarize an image by error diffusion - iion: copy named input to named output (useful to change file format) 2. Accumulators: combine several images into one - plambda: apply an arbitrary pixel-wise expression, given explicitly - veco: generic pixel-wise expressions for gray images, gray output - vecov: generic pixel-wise expressions for color images, color output - vecoh: generic pixel-wise operations for gray images, color output - tbcat: concatenate images top-bottom - lrcat: concatenate images left-right 3. Queries: extract data from one image - imprintf: print a formatted string of image data - ghisto: common histogram - contihist: continuous histogram - viewflow: represent a vector field using a color code - flowarrows: represent a vector field using arrows 4. Interpolators: fill-in the holes of one image - nnint: nearest neighbor interpolation - bdint: interpolation from the boundary of each hole - simpois: poisson, laplace and biharmonic interpolators - amle: absolutely minimizing lipschitz extension 5. Rescalers: produce an image of different size or shape - downsa: zoom out by combining blocks of pixels - upsa: zoom in by interpolation inside cells - ntiply: zoom in by pixel replication - imflip: rotate or transpose the image domain - homwarp: apply an arbitrary homography to the image domain 6. Frequecy domains: - fft: discrete Fourier transform (direct) - ifft: discrete Fourier transform (inverse) - dct: discrete cosine transform - dht: discrete Hartley transform 7. Point processing: - pview: display points or matches as an image - ransac: generic ransac implementation - srmatch: multi-scale sift matching for registration - plyflatten: project 3D points into a 2.5D representation - colormesh: build 3D mesh from calibrated 2.5D representation - ijmesh: build 3D mesh from non-calibrated 2.5D representation 8. Multi-program suites: - siftu: operations for sift keypoints - tiffu: operations for tiff files - fontu: operations for bitmapped fonts 9. Miscellaneous: - synflow: generate a synthetic optical flow field - ccproc: connected component filtering - ihough2: generic hough transform (for straight lines) - tdip: cylindrical hough transform (for sinusoids) - rpc_pm: patch match in the altitude domain - distance: distance function to a given set of points - sdistance: signed distance to the boundaries of a binary image - ppsmooth: periodic + smooth image decomposition - pmba2: poor man bundle adjustment - tcregistration: register two images by translation 10. Interactive programs: - fpan: display an image with panning, zooming, and contrast changes - fpantiff: like fpan, but understands image pyramids - rpcflip: display several calibrated satellite images - powerkill: fourier-domain band-pas filter editor - epiview: visualize the epipolar geometry between two images - viho: interactive homographic visualization - dosdo: look at an image and its fourier transform - icrop: interactive crop - vnav: visualize cylindrical images
Gaussian blur of different sizes
blur gaussian 1 A.png A_gaussian_1.png blur gaussian 3 A.png A_gaussian_3.png blur gaussian 8 A.png A_gaussian_8.png
Gaussian blur with different boundary conditions:
blur gaussian 5 -s A.png A_gaussian_5_s.png # symmetric boundary blur gaussian 5 -p A.png A_gaussian_5_p.png # periodic boundary blur gaussian 5 -z A.png A_gaussian_5_z.png # zero boundary
Blur with different kernels of the "same" width:
blur square 11 A.png A_s.png blur gaussian 5 A.png A_g.png blur laplace 5 A.png A_l.png blur disk 5 A.png A_d.png blur cauchy 2 A.png A_c.png blur inverse 1 -p A.png | qauto -i -p 1 - A_i.png
morsi cross: morphological operations with a cross structuring element:
morsi cross erosion A.png A_x_erosion.png morsi cross dilation A.png A_x_dilation.png morsi cross median A.png A_x_median.png morsi cross opening A.png A_x_opening.png morsi cross closing A.png A_x_closing.png morsi cross gradient A.png A_x_gradient.png morsi cross igradient A.png A_x_igradient.png morsi cross egradient A.png A_x_egradient.png morsi cross laplacian A.png | qeasy 40 -40 - A_x_laplacian.png morsi cross enhance A.png A_x_enhance.png morsi cross tophat A.png | qeasy 0 60 - A_x_tophat.png morsi cross bothat A.png | qeasy 0 60 - A_x_bothat.png morsi cross oscillation A.png | qeasy 0 60 - A_x_oscillation.png
morsi disk5: morphological operations with disk structuring element of radius 5 pixels:
morsi disk5 erosion A.png A_d5_erosion.png morsi disk5 dilation A.png A_d5_dilation.png morsi disk5 median A.png A_d5_median.png morsi disk5 opening A.png A_d5_opening.png morsi disk5 closing A.png A_d5_closing.png morsi disk5 gradient A.png A_d5_gradient.png morsi disk5 igradient A.png A_d5_igradient.png morsi disk5 egradient A.png A_d5_egradient.png morsi disk5 laplacian A.png | qeasy 90 -90 - A_d5_laplacian.png morsi disk5 enhance A.png A_d5_enhance.png morsi disk5 tophat A.png | qeasy 0 100 - A_d5_tophat.png morsi disk5 bothat A.png | qeasy 0 100 - A_d5_bothat.png morsi disk5 oscillation A.png | qeasy 0 100 - A_d5_oscillation.png
qeasy: linear contrast change with manually chosen parameters:
qeasy 20 200 A.png A_saturated.png qeasy 0 1000 A.png A_darkened.png qeasy -300 300 A.png A_bleached.png
qauto: linear contrast change with automatically chosen parameters:
qauto lenac.png lenac_qauto.png qauto -p 10 lenac.png lenac_qauto_p10.png qauto -i lenac.png lenac_qauto_i.png
palette: apply a palette to a scalar-valued image (the -l option is optional and writes the associated legend). The palette name can be a pre-defined palette or a gimp palette file:
qauto X.tif X.png palette 0 1000 dem X.tif X_dem1000.png -l leg_dem1000.png palette 0 2000 dem X.tif X_dem2000.png -l leg_dem2000.png palette 0 500 dem X.tif X_dem500.png -l leg_dem500.png palette nan nan gray X.tif X_gray.png -l leg_gray.png palette nan 1000 jet.gpl X.tif X_jet.png -l leg_jet.png palette 0 2000 nice X.tif X_nice.png -l leg_nice.png palette 0 2000 nnice X.tif X_nnice.png -l leg_nnice.png
dither: halftoning by error diffusion:
dither lena.png lena_dithered.png blur C 1 lena.png | qauto | dither - lena_dithered_c.png
plambda: apply an arithmetic expression with input images (the
expression is written in reverse polish
notation). Common arithmetic operators and all functions from
math.h are allowed.
plambda a.jpg b.jpg + -o ab_sum.png plambda a.jpg b.jpg - -o ab_dif.png plambda a.jpg b.jpg "*" -o ab_mul.png plambda a.jpg b.jpg / -o ab_div.png plambda a.jpg b.jpg "+ 2 /" -o ab_avg.png plambda a.jpg b.jpg "* sqrt" -o ab_geo.png plambda a.jpg b.jpg fmin -o ab_min.png plambda a.jpg b.jpg fmax -o ab_max.png
The plambda language allows many operations, even when applied to a single image. See the external plambda tutorial for a comprehensive description. Here are some silly examples:
plambda b.jpg "a[2] a[1] a[0] join3" -o b_bgr.png # swap colors plambda b.jpg "a[0] a[1] a[2] join3" -o b_rgb.png # do nothing plambda b.jpg "randg 20 * +" -o b_noisy.png # add luminance noise plambda b.jpg "a(1,0) a -" -o b_dx.png # x-derivative plambda b.jpg "a(1,0) a - 127 +" -o b_dx_127.png # x-derivative plus 127 plambda b.jpg a,x -o b_dx_alt.png # x-derivative (easier) plambda b.jpg "a,x a,y hypot 2 *" -o b_edges.png # gradient norm plambda b.jpg ":y 50 * sin 50 * +" -o b_bands.png # sine of y plus image plambda b.jpg ":r 50 * sin 50 * +" -o b_rings.png # sine of r plus image plambda b.jpg ":i 8 fmod not :j 30 fmod not or 99 * +" -o b_grid.png
The lrcat and tbcat programs concatenate images horizontally or vertically. The images need not be of the same size. Any missing space is filled-in by the value on the environment variable BACKGROUND.
lrcat a.jpg b.jpg -o ab_lr.png tbcat a.jpg b.jpg -o ab_tb.png lrcat a.jpg b.jpg | tbcat - lena.png -o ab_montage.png lrcat a.jpg b.jpg | BACKGROUND=255 tbcat - lena.png -o ab_montage2.png
The vecov program aggregates the pixel values of many images.
vecov avg pRi*.jpg -o outv_avg.png vecov min pRi*.jpg -o outv_min.png vecov max pRi*.jpg -o outv_max.png vecov med pRi*.jpg -o outv_med.png # geometric medoid vecov modc pRi*.jpg -o outv_modc.png # component-wise mode
The veco program is similar, but it only works for gray-scale images and supports a much larger variety of accumulators (including very fancy ones).
veco max gpRi*.jpg -o out_max.png veco euc gpRi*.jpg -o out_euc.png # euclidean mean veco avg gpRi*.jpg -o out_avg.png # arithmetic mean veco geo gpRi*.jpg -o out_geo.png # geometric mean veco har gpRi*.jpg -o out_har.png # harmonic mean veco min gpRi*.jpg -o out_min.png veco lav gpRi*.jpg -o out_lav.png # logarithmic average veco lse gpRi*.jpg -o out_lse.png # log-sum-exp (soft max) veco mod gpRi*.jpg -o out_mod.png # mode veco med gpRi*.jpg -o out_med.png # median veco q25 gpRi*.jpg -o out_q25.png # first quartile veco q75 gpRi*.jpg -o out_q75.png # third quartile veco rnd gpRi*.jpg -o out_rnd.png # random sample veco std gpRi*.jpg | qeasy 0 20 - out_std.png # standard deviation veco iqd gpRi*.jpg | qeasy 0 20 - out_iqd.png # interquartile distance
The vecoh program also works with gray-scale images and it finds clusters of values at each pixel. The output is a 8-dimesional image, on the first component the number of clusters, and on the rest, the cluster centers.
vecoh kmeans gpRi*jpg | plambda - "x[0] 0 4 qe" -o kmeans_count.png vecoh kmedians gpRi*jpg | plambda - "x[0] 0 4 qe" -o kmedians_count.png vecoh contrario gpRi*jpg | plambda - "x[0] 0 4 qe" -o contrario_count.png vecoh kmeans gpRi*jpg | plambda - "x[1]" -o kmeans_k1.png vecoh kmeans gpRi*jpg | plambda - "x[2]" -o kmeans_k2.png vecoh kmeans gpRi*jpg | plambda - "x[3]" -o kmeans_k3.png vecoh kmeans gpRi*jpg | plambda - "x[4]" -o kmeans_k4.png vecoh kmedians gpRi*jpg | plambda - "x[1]" -o kmedians_k1.png vecoh kmedians gpRi*jpg | plambda - "x[2]" -o kmedians_k2.png vecoh kmedians gpRi*jpg | plambda - "x[3]" -o kmedians_k3.png vecoh kmedians gpRi*jpg | plambda - "x[4]" -o kmedians_k4.png vecoh contrario gpRi*jpg | plambda - "x[1]" -o contrario_k1.png vecoh contrario gpRi*jpg | plambda - "x[2]" -o contrario_k2.png
The imprintf program creates a string from image data. The syntax is very similar to C's printf, with a image-specific set of conversions.
imprintf "Image size=%wx%h avg=%v med=%e min=%i max=%a\n" lena.png
Image size=256x256 avg=119.08 med=119.08 min=0 max=255
The ghisto program computes the histogram of an image and prints it in a
textual form that gnuplot understands. The -p option produces a
gnuplot file that renders into a png.
plambda lena.png "40 / round 40 *" | ghisto
set xrange [0:240] set yrange [0:] set format y "" unset key plot "-" w impulses title "histogram", "-" w lines title "accumulated histogram" 0 5903 40 7993 80 10559 120 13691 160 16375 200 6243 240 4772 end 0 1474.94 40 3472.09 80 6110.39 120 9531.26 160 13622.8 200 15182.7 240 16375 end
cat lena.png | ghisto -p | gnuplot > lena_ghisto.png plambda lena.png "10 / round 10 *" | ghisto -p | gnuplot > q10lena_ghisto.png plambda lena.png "40 / round 40 *" | ghisto -p | gnuplot > q40lena_ghisto.png blur g 0.4 lena.png | ghisto -p | gnuplot > glena_ghisto.png
The contihist program computes the continuous histogram of an image (defined as the derivative of the area of level sets of the interpolated image). This is a real-valued function of a real variable, and you have to specify the number of samples to represent this function function and the desired range.
cat lena.png | contihist 100 0 255 - -p | gnuplot > lena_contih_100.png cat lena.png | contihist 1000 0 255 - -p | gnuplot > lena_contih_1000.png cat lena.png | contihist 2000 0 255 - -p | gnuplot > lena_contih_2000.png cat lena.png | contihist 10000 0 255 - -p | gnuplot > lena_contih_10000.png
Images with two-dimensional pixels are very important in image processing. They can be for example displacement fields, velocity fields, gradients of gray-valued images, complex-valued images (such as SLC radar images, or Fourier transforms of gray-valued images). Depending on the context, several different visualizations are preferred.
cat wheel.tiff | viewflow 0 - wheel_colors.png cat wheel.tiff | viewflow -1 - wheel_lines.png cat wheel.tiff | flowarrows 0.2 17 - wheel_arrows.png cat wheel.tiff | plambda x[0] | qeasy -1 1 - wheel_x.png cat wheel.tiff | plambda x[1] | qeasy -1 1 - wheel_y.png cat wheel.tiff | plambda vnorm | qeasy 0 2 - wheel_abs.png cat wheel.tiff | plambda "x[1] x[0] atan2" | qeasy -3.14 3.14 - wheel_arg.png cat wheel.tiff | plambda "split dup join3" | qeasy -2 2 - wheel_xyy.png
plambda lena.png x,g | viewflow 0 - g_colors.png plambda lena.png x,g | viewflow -50 - g_lines.png plambda lena.png x,g | flowarrows 3 5 - g_arrows.png plambda lena.png x,g | plambda x[0] | qauto - g_x.png plambda lena.png x,g | plambda x[1] | qauto - g_y.png plambda lena.png x,g | plambda vnorm | qauto - g_abs.png plambda lena.png x,g | plambda "x[1] x[0] atan2" | qauto - g_arg.png plambda lena.png x,g | plambda "split dup join3" | qauto - g_xyy.png
cat slc.tiff | viewflow 0 - slc_colors.png cat slc.tiff | viewflow -500 - slc_lines.png cat slc.tiff | flowarrows 3 5 - slc_arrows.png cat slc.tiff | plambda x[0] | qauto - slc_x.png cat slc.tiff | plambda x[1] | qauto - slc_y.png cat slc.tiff | plambda vnorm | qauto - slc_abs.png -p 0.1 cat slc.tiff | plambda "x[1] x[0] atan2" | qauto - slc_arg.png cat slc.tiff | plambda "split dup join3" | qauto - slc_xyy.png
These interpolating filters are used to fill-in the missing values of an image (by default, indicated by a floating point NAN).
qeasy 0 255 masked.tif masked.png cat masked.tif | nnint - out_nnint.png # nearest-neighbor cat masked.tif | bdint -a min - out_bdmin.png # boundary minimum cat masked.tif | bdint -a max - out_bdmax.png # boundary maximum cat masked.tif | bdint -a avg - out_bdavg.png # boundary average cat masked.tif | simpois -o out_laplace.png # harmonic cat masked.tif | simpois -t -0.08 -o out_biharm.png # biharmonic
Zoom-in is a particular case of interpolation when the data values are on a regular grid. Zoom-out is a very different problem, typically achieved by filtering the input image and then sampling it at a coarser grid.
The downsa program creates a smaller image by combining blocks of n×n pixels into one. The first argument is the rule to combine several pixel values into one. The rule is specified by a single letter:
downsa v 2 masked.png downsa_avg_2.png # zoom-out by 2x2 blocks aVerage downsa i 2 masked.png downsa_min_2.png # zoom-out by 2x2 blocks mInimum downsa a 2 masked.png downsa_max_2.png # zoom-out by 2x2 blocks mAximum downsa e 2 masked.png downsa_med_2.png # zoom-out by 2x2 blocks mEdian downsa f 2 masked.png downsa_1st_2.png # zoom-out by 2x2 blocks First pixel downsa l 2 masked.png downsa_lst_2.png # zoom-out by 2x2 blocks Last pixel downsa r 2 masked.png downsa_rnd_2.png # zoom-out by 2x2 blocks Random pixel
The second argument to the downsa program is the zoom factor, always a positive integer:
downsa a 9 masked.png downsa_max_9.png downsa a 8 masked.png downsa_max_8.png downsa a 7 masked.png downsa_max_7.png downsa a 6 masked.png downsa_max_6.png downsa a 5 masked.png downsa_max_5.png downsa a 4 masked.png downsa_max_4.png downsa a 3 masked.png downsa_max_3.png downsa a 2 masked.png downsa_max_2.png downsa a 1 masked.png downsa_max_1.png
The ntiply program is an inverse of downsa: it creates a larger image by replicating each pixel into a block of n×n. The dimensions of the images are multiplied by n, exactly.
downsa a 9 masked.png | ntiply 9 - ntiply_9.png downsa a 8 masked.png | ntiply 8 - ntiply_8.png downsa a 7 masked.png | ntiply 7 - ntiply_7.png downsa a 6 masked.png | ntiply 6 - ntiply_6.png downsa a 5 masked.png | ntiply 5 - ntiply_5.png downsa a 4 masked.png | ntiply 4 - ntiply_4.png downsa a 3 masked.png | ntiply 3 - ntiply_3.png downsa a 2 masked.png | ntiply 2 - ntiply_2.png downsa a 1 masked.png | ntiply 1 - ntiply_1.png
The upsa program is another inverse of downsa: it creates a larger image by interpolating the pixel values on a grid with n×n the resolution. Notice that an image of size W&ntimes;H is transformed into an image of size (nW-n)×(nH-n), because the interpolator needs values at both sides of the new samples. The second argument is the order of the spline used to interpolate:
downsa v 7 A.png A_downsa7.png downsa v 7 A.png | upsa 7 0 - upsa_7_0.png # nearest-neighbor interpolation downsa v 7 A.png | upsa 7 1 - upsa_7_1.png # piecewise linear downsa v 7 A.png | upsa 7 2 - upsa_7_2.png # bilinear downsa v 7 A.png | upsa 7 3 - upsa_7_3.png # bicubic downsa v 7 A.png | upsa 7 -2 - upsa_7_-2.png # bilinear with Quilez fading downsa v 7 A.png | upsa 7 -3 - upsa_7_-3.png # bicubic with Quilez fading
Notice that upsa with nearest-neighbor interpolation looks similar to ntiply, but it is not identical (the images have different size).
downsa v 20 b.jpg | upsa 20 0 - b_upsa_20.png downsa v 20 b.jpg | ntiply 20 - b_ntiply_20.png downsa v 20 b.jpg|upsa 20 0|GETPIXEL=0 plambda zero:240x240 - "x y(-10,-10)" -o b_upsabg_20.png
The program homwarp allows a finer control of the transformation: you get to specify the 9 coefficients of an arbitrary homography and the desired size of the output image. Notice that this generalizes upsa and ntiply but it does much more: it allows for arbitrary rotations and scalings.
A very common particular case of rotations are those of straight angles. They are directly performed by the program imflip:
imflip r90 b.jpg b_r90.png imflip r180 b.jpg b_r180.png imflip r270 b.jpg b_r270.png imflip leftright b.jpg b_leftright.png imflip topdown b.jpg b_topdown.png imflip transpose b.jpg b_transpose.png imflip posetrans b.jpg b_posetrans.png imflip identity b.jpg b_identity.png
The fft, ifft, dct and dht programs are filters tha compute, respectively, the discrete Fourier transform, its inverse, the discrete cosine transform and the discrete Hartley transform. The DCT and the DHT are self-inverses, and the IFFT is the inverse of the FFT.
cat barb.png | fft | plambda vnorm | qauto -p 1 - barb_dft_abs.png cat barb.png | fft | plambda "x[0]" | qauto -p 1 - barb_dft_re.png cat barb.png | fft | plambda "x[1]" | qauto -p 1 - barb_dft_im.png cat barb.png | fft | plambda "x[0] fabs" | qauto -p 1 - barb_dft_are.png cat barb.png | fft | plambda "x[1] fabs" | qauto -p 1 - barb_dft_aim.png cat barb.png | dht | qauto -p 1 - barb_dht.png cat barb.png | dht | plambda fabs | qauto -p 1 - barb_dht_abs.png cat barb.png | dct | qauto -p 1 - barb_dct.png cat barb.png | dct | plambda fabs | qauto -p 1 - barb_dct_abs.png
fftshift barb.png barb_s.png cat barb.png | fft | plambda vnorm | fftshift | qauto -p 1 - barb_dft_s.png cat barb.png | dht | plambda vnorm | fftshift | qauto -p 1 - barb_dht_abs_s.png cat barb.png | dht | fftshift | qauto -p 1 - barb_dht_s.png
end
