[PATCH v3 03/23] libtuning: Copy files from raspberrypi
Laurent Pinchart
laurent.pinchart at ideasonboard.com
Thu Jul 4 11:20:39 CEST 2024
On Wed, Jul 03, 2024 at 04:16:52PM +0200, Stefan Klug wrote:
> Copy ctt_{awb,ccm,colors,ransac} from the raspberrypi tuning scripts as
> basis for the libcamera implementation. color.py was renamed to
> ctt_colors.py to better express the origin.
>
> The files were taken from commit 66479605baca4a22e2b
The files were taken from commit 66479605baca ("utils: raspberrypi: ctt:
Improve the Macbeth Chart search reliability").
> Signed-off-by: Stefan Klug <stefan.klug at ideasonboard.com>
> Acked-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
Acked-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
> ---
> utils/tuning/libtuning/ctt_awb.py | 376 +++++++++++++++++++++++++
> utils/tuning/libtuning/ctt_ccm.py | 406 +++++++++++++++++++++++++++
> utils/tuning/libtuning/ctt_colors.py | 30 ++
> utils/tuning/libtuning/ctt_ransac.py | 71 +++++
> 4 files changed, 883 insertions(+)
> create mode 100644 utils/tuning/libtuning/ctt_awb.py
> create mode 100644 utils/tuning/libtuning/ctt_ccm.py
> create mode 100644 utils/tuning/libtuning/ctt_colors.py
> create mode 100644 utils/tuning/libtuning/ctt_ransac.py
>
> diff --git a/utils/tuning/libtuning/ctt_awb.py b/utils/tuning/libtuning/ctt_awb.py
> new file mode 100644
> index 000000000000..5ba6f978a228
> --- /dev/null
> +++ b/utils/tuning/libtuning/ctt_awb.py
> @@ -0,0 +1,376 @@
> +# SPDX-License-Identifier: BSD-2-Clause
> +#
> +# Copyright (C) 2019, Raspberry Pi Ltd
> +#
> +# camera tuning tool for AWB
> +
> +from ctt_image_load import *
> +import matplotlib.pyplot as plt
> +from bisect import bisect_left
> +from scipy.optimize import fmin
> +
> +
> +"""
> +obtain piecewise linear approximation for colour curve
> +"""
> +def awb(Cam, cal_cr_list, cal_cb_list, plot):
> + imgs = Cam.imgs
> + """
> + condense alsc calibration tables into one dictionary
> + """
> + if cal_cr_list is None:
> + colour_cals = None
> + else:
> + colour_cals = {}
> + for cr, cb in zip(cal_cr_list, cal_cb_list):
> + cr_tab = cr['table']
> + cb_tab = cb['table']
> + """
> + normalise tables so min value is 1
> + """
> + cr_tab = cr_tab/np.min(cr_tab)
> + cb_tab = cb_tab/np.min(cb_tab)
> + colour_cals[cr['ct']] = [cr_tab, cb_tab]
> + """
> + obtain data from greyscale macbeth patches
> + """
> + rb_raw = []
> + rbs_hat = []
> + for Img in imgs:
> + Cam.log += '\nProcessing '+Img.name
> + """
> + get greyscale patches with alsc applied if alsc enabled.
> + Note: if alsc is disabled then colour_cals will be set to None and the
> + function will just return the greyscale patches
> + """
> + r_patchs, b_patchs, g_patchs = get_alsc_patches(Img, colour_cals)
> + """
> + calculate ratio of r, b to g
> + """
> + r_g = np.mean(r_patchs/g_patchs)
> + b_g = np.mean(b_patchs/g_patchs)
> + Cam.log += '\n r : {:.4f} b : {:.4f}'.format(r_g, b_g)
> + """
> + The curve tends to be better behaved in so-called hatspace.
> + R, B, G represent the individual channels. The colour curve is plotted in
> + r, b space, where:
> + r = R/G
> + b = B/G
> + This will be referred to as dehatspace... (sorry)
> + Hatspace is defined as:
> + r_hat = R/(R+B+G)
> + b_hat = B/(R+B+G)
> + To convert from dehatspace to hastpace (hat operation):
> + r_hat = r/(1+r+b)
> + b_hat = b/(1+r+b)
> + To convert from hatspace to dehatspace (dehat operation):
> + r = r_hat/(1-r_hat-b_hat)
> + b = b_hat/(1-r_hat-b_hat)
> + Proof is left as an excercise to the reader...
> + Throughout the code, r and b are sometimes referred to as r_g and b_g
> + as a reminder that they are ratios
> + """
> + r_g_hat = r_g/(1+r_g+b_g)
> + b_g_hat = b_g/(1+r_g+b_g)
> + Cam.log += '\n r_hat : {:.4f} b_hat : {:.4f}'.format(r_g_hat, b_g_hat)
> + rbs_hat.append((r_g_hat, b_g_hat, Img.col))
> + rb_raw.append((r_g, b_g))
> + Cam.log += '\n'
> +
> + Cam.log += '\nFinished processing images'
> + """
> + sort all lits simultaneously by r_hat
> + """
> + rbs_zip = list(zip(rbs_hat, rb_raw))
> + rbs_zip.sort(key=lambda x: x[0][0])
> + rbs_hat, rb_raw = list(zip(*rbs_zip))
> + """
> + unzip tuples ready for processing
> + """
> + rbs_hat = list(zip(*rbs_hat))
> + rb_raw = list(zip(*rb_raw))
> + """
> + fit quadratic fit to r_g hat and b_g_hat
> + """
> + a, b, c = np.polyfit(rbs_hat[0], rbs_hat[1], 2)
> + Cam.log += '\nFit quadratic curve in hatspace'
> + """
> + the algorithm now approximates the shortest distance from each point to the
> + curve in dehatspace. Since the fit is done in hatspace, it is easier to
> + find the actual shortest distance in hatspace and use the projection back
> + into dehatspace as an overestimate.
> + The distance will be used for two things:
> + 1) In the case that colour temperature does not strictly decrease with
> + increasing r/g, the closest point to the line will be chosen out of an
> + increasing pair of colours.
> +
> + 2) To calculate transverse negative an dpositive, the maximum positive
> + and negative distance from the line are chosen. This benefits from the
> + overestimate as the transverse pos/neg are upper bound values.
> + """
> + """
> + define fit function
> + """
> + def f(x):
> + return a*x**2 + b*x + c
> + """
> + iterate over points (R, B are x and y coordinates of points) and calculate
> + distance to line in dehatspace
> + """
> + dists = []
> + for i, (R, B) in enumerate(zip(rbs_hat[0], rbs_hat[1])):
> + """
> + define function to minimise as square distance between datapoint and
> + point on curve. Squaring is monotonic so minimising radius squared is
> + equivalent to minimising radius
> + """
> + def f_min(x):
> + y = f(x)
> + return((x-R)**2+(y-B)**2)
> + """
> + perform optimisation with scipy.optmisie.fmin
> + """
> + x_hat = fmin(f_min, R, disp=0)[0]
> + y_hat = f(x_hat)
> + """
> + dehat
> + """
> + x = x_hat/(1-x_hat-y_hat)
> + y = y_hat/(1-x_hat-y_hat)
> + rr = R/(1-R-B)
> + bb = B/(1-R-B)
> + """
> + calculate euclidean distance in dehatspace
> + """
> + dist = ((x-rr)**2+(y-bb)**2)**0.5
> + """
> + return negative if point is below the fit curve
> + """
> + if (x+y) > (rr+bb):
> + dist *= -1
> + dists.append(dist)
> + Cam.log += '\nFound closest point on fit line to each point in dehatspace'
> + """
> + calculate wiggle factors in awb. 10% added since this is an upper bound
> + """
> + transverse_neg = - np.min(dists) * 1.1
> + transverse_pos = np.max(dists) * 1.1
> + Cam.log += '\nTransverse pos : {:.5f}'.format(transverse_pos)
> + Cam.log += '\nTransverse neg : {:.5f}'.format(transverse_neg)
> + """
> + set minimum transverse wiggles to 0.1 .
> + Wiggle factors dictate how far off of the curve the algorithm searches. 0.1
> + is a suitable minimum that gives better results for lighting conditions not
> + within calibration dataset. Anything less will generalise poorly.
> + """
> + if transverse_pos < 0.01:
> + transverse_pos = 0.01
> + Cam.log += '\nForced transverse pos to 0.01'
> + if transverse_neg < 0.01:
> + transverse_neg = 0.01
> + Cam.log += '\nForced transverse neg to 0.01'
> +
> + """
> + generate new b_hat values at each r_hat according to fit
> + """
> + r_hat_fit = np.array(rbs_hat[0])
> + b_hat_fit = a*r_hat_fit**2 + b*r_hat_fit + c
> + """
> + transform from hatspace to dehatspace
> + """
> + r_fit = r_hat_fit/(1-r_hat_fit-b_hat_fit)
> + b_fit = b_hat_fit/(1-r_hat_fit-b_hat_fit)
> + c_fit = np.round(rbs_hat[2], 0)
> + """
> + round to 4dp
> + """
> + r_fit = np.where((1000*r_fit) % 1 <= 0.05, r_fit+0.0001, r_fit)
> + r_fit = np.where((1000*r_fit) % 1 >= 0.95, r_fit-0.0001, r_fit)
> + b_fit = np.where((1000*b_fit) % 1 <= 0.05, b_fit+0.0001, b_fit)
> + b_fit = np.where((1000*b_fit) % 1 >= 0.95, b_fit-0.0001, b_fit)
> + r_fit = np.round(r_fit, 4)
> + b_fit = np.round(b_fit, 4)
> + """
> + The following code ensures that colour temperature decreases with
> + increasing r/g
> + """
> + """
> + iterate backwards over list for easier indexing
> + """
> + i = len(c_fit) - 1
> + while i > 0:
> + if c_fit[i] > c_fit[i-1]:
> + Cam.log += '\nColour temperature increase found\n'
> + Cam.log += '{} K at r = {} to '.format(c_fit[i-1], r_fit[i-1])
> + Cam.log += '{} K at r = {}'.format(c_fit[i], r_fit[i])
> + """
> + if colour temperature increases then discard point furthest from
> + the transformed fit (dehatspace)
> + """
> + error_1 = abs(dists[i-1])
> + error_2 = abs(dists[i])
> + Cam.log += '\nDistances from fit:\n'
> + Cam.log += '{} K : {:.5f} , '.format(c_fit[i], error_1)
> + Cam.log += '{} K : {:.5f}'.format(c_fit[i-1], error_2)
> + """
> + find bad index
> + note that in python false = 0 and true = 1
> + """
> + bad = i - (error_1 < error_2)
> + Cam.log += '\nPoint at {} K deleted as '.format(c_fit[bad])
> + Cam.log += 'it is furthest from fit'
> + """
> + delete bad point
> + """
> + r_fit = np.delete(r_fit, bad)
> + b_fit = np.delete(b_fit, bad)
> + c_fit = np.delete(c_fit, bad).astype(np.uint16)
> + """
> + note that if a point has been discarded then the length has decreased
> + by one, meaning that decreasing the index by one will reassess the kept
> + point against the next point. It is therefore possible, in theory, for
> + two adjacent points to be discarded, although probably rare
> + """
> + i -= 1
> +
> + """
> + return formatted ct curve, ordered by increasing colour temperature
> + """
> + ct_curve = list(np.array(list(zip(b_fit, r_fit, c_fit))).flatten())[::-1]
> + Cam.log += '\nFinal CT curve:'
> + for i in range(len(ct_curve)//3):
> + j = 3*i
> + Cam.log += '\n ct: {} '.format(ct_curve[j])
> + Cam.log += ' r: {} '.format(ct_curve[j+1])
> + Cam.log += ' b: {} '.format(ct_curve[j+2])
> +
> + """
> + plotting code for debug
> + """
> + if plot:
> + x = np.linspace(np.min(rbs_hat[0]), np.max(rbs_hat[0]), 100)
> + y = a*x**2 + b*x + c
> + plt.subplot(2, 1, 1)
> + plt.title('hatspace')
> + plt.plot(rbs_hat[0], rbs_hat[1], ls='--', color='blue')
> + plt.plot(x, y, color='green', ls='-')
> + plt.scatter(rbs_hat[0], rbs_hat[1], color='red')
> + for i, ct in enumerate(rbs_hat[2]):
> + plt.annotate(str(ct), (rbs_hat[0][i], rbs_hat[1][i]))
> + plt.xlabel('$\\hat{r}$')
> + plt.ylabel('$\\hat{b}$')
> + """
> + optional set axes equal to shortest distance so line really does
> + looks perpendicular and everybody is happy
> + """
> + # ax = plt.gca()
> + # ax.set_aspect('equal')
> + plt.grid()
> + plt.subplot(2, 1, 2)
> + plt.title('dehatspace - indoors?')
> + plt.plot(r_fit, b_fit, color='blue')
> + plt.scatter(rb_raw[0], rb_raw[1], color='green')
> + plt.scatter(r_fit, b_fit, color='red')
> + for i, ct in enumerate(c_fit):
> + plt.annotate(str(ct), (r_fit[i], b_fit[i]))
> + plt.xlabel('$r$')
> + plt.ylabel('$b$')
> + """
> + optional set axes equal to shortest distance so line really does
> + looks perpendicular and everybody is happy
> + """
> + # ax = plt.gca()
> + # ax.set_aspect('equal')
> + plt.subplots_adjust(hspace=0.5)
> + plt.grid()
> + plt.show()
> + """
> + end of plotting code
> + """
> + return(ct_curve, np.round(transverse_pos, 5), np.round(transverse_neg, 5))
> +
> +
> +"""
> +obtain greyscale patches and perform alsc colour correction
> +"""
> +def get_alsc_patches(Img, colour_cals, grey=True):
> + """
> + get patch centre coordinates, image colour and the actual
> + patches for each channel, remembering to subtract blacklevel
> + If grey then only greyscale patches considered
> + """
> + if grey:
> + cen_coords = Img.cen_coords[3::4]
> + col = Img.col
> + patches = [np.array(Img.patches[i]) for i in Img.order]
> + r_patchs = patches[0][3::4] - Img.blacklevel_16
> + b_patchs = patches[3][3::4] - Img.blacklevel_16
> + """
> + note two green channels are averages
> + """
> + g_patchs = (patches[1][3::4]+patches[2][3::4])/2 - Img.blacklevel_16
> + else:
> + cen_coords = Img.cen_coords
> + col = Img.col
> + patches = [np.array(Img.patches[i]) for i in Img.order]
> + r_patchs = patches[0] - Img.blacklevel_16
> + b_patchs = patches[3] - Img.blacklevel_16
> + g_patchs = (patches[1]+patches[2])/2 - Img.blacklevel_16
> +
> + if colour_cals is None:
> + return r_patchs, b_patchs, g_patchs
> + """
> + find where image colour fits in alsc colour calibration tables
> + """
> + cts = list(colour_cals.keys())
> + pos = bisect_left(cts, col)
> + """
> + if img colour is below minimum or above maximum alsc calibration colour, simply
> + pick extreme closest to img colour
> + """
> + if pos % len(cts) == 0:
> + """
> + this works because -0 = 0 = first and -1 = last index
> + """
> + col_tabs = np.array(colour_cals[cts[-pos//len(cts)]])
> + """
> + else, perform linear interpolation between existing alsc colour
> + calibration tables
> + """
> + else:
> + bef = cts[pos-1]
> + aft = cts[pos]
> + da = col-bef
> + db = aft-col
> + bef_tabs = np.array(colour_cals[bef])
> + aft_tabs = np.array(colour_cals[aft])
> + col_tabs = (bef_tabs*db + aft_tabs*da)/(da+db)
> + col_tabs = np.reshape(col_tabs, (2, 12, 16))
> + """
> + calculate dx, dy used to calculate alsc table
> + """
> + w, h = Img.w/2, Img.h/2
> + dx, dy = int(-(-(w-1)//16)), int(-(-(h-1)//12))
> + """
> + make list of pairs of gains for each patch by selecting the correct value
> + in alsc colour calibration table
> + """
> + patch_gains = []
> + for cen in cen_coords:
> + x, y = cen[0]//dx, cen[1]//dy
> + # We could probably do with some better spatial interpolation here?
> + col_gains = (col_tabs[0][y][x], col_tabs[1][y][x])
> + patch_gains.append(col_gains)
> +
> + """
> + multiply the r and b channels in each patch by the respective gain, finally
> + performing the alsc colour correction
> + """
> + for i, gains in enumerate(patch_gains):
> + r_patchs[i] = r_patchs[i] * gains[0]
> + b_patchs[i] = b_patchs[i] * gains[1]
> +
> + """
> + return greyscale patches, g channel and correct r, b channels
> + """
> + return r_patchs, b_patchs, g_patchs
> diff --git a/utils/tuning/libtuning/ctt_ccm.py b/utils/tuning/libtuning/ctt_ccm.py
> new file mode 100644
> index 000000000000..59753e332ee9
> --- /dev/null
> +++ b/utils/tuning/libtuning/ctt_ccm.py
> @@ -0,0 +1,406 @@
> +# SPDX-License-Identifier: BSD-2-Clause
> +#
> +# Copyright (C) 2019, Raspberry Pi Ltd
> +#
> +# camera tuning tool for CCM (colour correction matrix)
> +
> +from ctt_image_load import *
> +from ctt_awb import get_alsc_patches
> +import colors
> +from scipy.optimize import minimize
> +from ctt_visualise import visualise_macbeth_chart
> +import numpy as np
> +"""
> +takes 8-bit macbeth chart values, degammas and returns 16 bit
> +"""
> +
> +'''
> +This program has many options from which to derive the color matrix from.
> +The first is average. This minimises the average delta E across all patches of
> +the macbeth chart. Testing across all cameras yeilded this as the most color
> +accurate and vivid. Other options are avalible however.
> +Maximum minimises the maximum Delta E of the patches. It iterates through till
> +a minimum maximum is found (so that there is
> +not one patch that deviates wildly.)
> +This yields generally good results but overall the colors are less accurate
> +Have a fiddle with maximum and see what you think.
> +The final option allows you to select the patches for which to average across.
> +This means that you can bias certain patches, for instance if you want the
> +reds to be more accurate.
> +'''
> +
> +matrix_selection_types = ["average", "maximum", "patches"]
> +typenum = 0 # select from array above, 0 = average, 1 = maximum, 2 = patches
> +test_patches = [1, 2, 5, 8, 9, 12, 14]
> +
> +'''
> +Enter patches to test for. Can also be entered twice if you
> +would like twice as much bias on one patch.
> +'''
> +
> +
> +def degamma(x):
> + x = x / ((2 ** 8) - 1) # takes 255 and scales it down to one
> + x = np.where(x < 0.04045, x / 12.92, ((x + 0.055) / 1.055) ** 2.4)
> + x = x * ((2 ** 16) - 1) # takes one and scales up to 65535, 16 bit color
> + return x
> +
> +
> +def gamma(x):
> + # Take 3 long array of color values and gamma them
> + return [((colour / 255) ** (1 / 2.4) * 1.055 - 0.055) * 255 for colour in x]
> +
> +
> +"""
> +FInds colour correction matrices for list of images
> +"""
> +
> +
> +def ccm(Cam, cal_cr_list, cal_cb_list):
> + global matrix_selection_types, typenum
> + imgs = Cam.imgs
> + """
> + standard macbeth chart colour values
> + """
> + m_rgb = np.array([ # these are in RGB
> + [116, 81, 67], # dark skin
> + [199, 147, 129], # light skin
> + [91, 122, 156], # blue sky
> + [90, 108, 64], # foliage
> + [130, 128, 176], # blue flower
> + [92, 190, 172], # bluish green
> + [224, 124, 47], # orange
> + [68, 91, 170], # purplish blue
> + [198, 82, 97], # moderate red
> + [94, 58, 106], # purple
> + [159, 189, 63], # yellow green
> + [230, 162, 39], # orange yellow
> + [35, 63, 147], # blue
> + [67, 149, 74], # green
> + [180, 49, 57], # red
> + [238, 198, 20], # yellow
> + [193, 84, 151], # magenta
> + [0, 136, 170], # cyan (goes out of gamut)
> + [245, 245, 243], # white 9.5
> + [200, 202, 202], # neutral 8
> + [161, 163, 163], # neutral 6.5
> + [121, 121, 122], # neutral 5
> + [82, 84, 86], # neutral 3.5
> + [49, 49, 51] # black 2
> + ])
> + """
> + convert reference colours from srgb to rgb
> + """
> + m_srgb = degamma(m_rgb) # now in 16 bit color.
> +
> + # Produce array of LAB values for ideal color chart
> + m_lab = [colors.RGB_to_LAB(color / 256) for color in m_srgb]
> +
> + """
> + reorder reference values to match how patches are ordered
> + """
> + m_srgb = np.array([m_srgb[i::6] for i in range(6)]).reshape((24, 3))
> + m_lab = np.array([m_lab[i::6] for i in range(6)]).reshape((24, 3))
> + m_rgb = np.array([m_rgb[i::6] for i in range(6)]).reshape((24, 3))
> + """
> + reformat alsc correction tables or set colour_cals to None if alsc is
> + deactivated
> + """
> + if cal_cr_list is None:
> + colour_cals = None
> + else:
> + colour_cals = {}
> + for cr, cb in zip(cal_cr_list, cal_cb_list):
> + cr_tab = cr['table']
> + cb_tab = cb['table']
> + """
> + normalise tables so min value is 1
> + """
> + cr_tab = cr_tab / np.min(cr_tab)
> + cb_tab = cb_tab / np.min(cb_tab)
> + colour_cals[cr['ct']] = [cr_tab, cb_tab]
> +
> + """
> + for each image, perform awb and alsc corrections.
> + Then calculate the colour correction matrix for that image, recording the
> + ccm and the colour tempertaure.
> + """
> + ccm_tab = {}
> + for Img in imgs:
> + Cam.log += '\nProcessing image: ' + Img.name
> + """
> + get macbeth patches with alsc applied if alsc enabled.
> + Note: if alsc is disabled then colour_cals will be set to None and no
> + the function will simply return the macbeth patches
> + """
> + r, b, g = get_alsc_patches(Img, colour_cals, grey=False)
> + # 256 values for each patch of sRGB values
> +
> + """
> + do awb
> + Note: awb is done by measuring the macbeth chart in the image, rather
> + than from the awb calibration. This is done so the awb will be perfect
> + and the ccm matrices will be more accurate.
> + """
> + r_greys, b_greys, g_greys = r[3::4], b[3::4], g[3::4]
> + r_g = np.mean(r_greys / g_greys)
> + b_g = np.mean(b_greys / g_greys)
> + r = r / r_g
> + b = b / b_g
> + """
> + normalise brightness wrt reference macbeth colours and then average
> + each channel for each patch
> + """
> + gain = np.mean(m_srgb) / np.mean((r, g, b))
> + Cam.log += '\nGain with respect to standard colours: {:.3f}'.format(gain)
> + r = np.mean(gain * r, axis=1)
> + b = np.mean(gain * b, axis=1)
> + g = np.mean(gain * g, axis=1)
> + """
> + calculate ccm matrix
> + """
> + # ==== All of below should in sRGB ===##
> + sumde = 0
> + ccm = do_ccm(r, g, b, m_srgb)
> + # This is the initial guess that our optimisation code works with.
> + original_ccm = ccm
> + r1 = ccm[0]
> + r2 = ccm[1]
> + g1 = ccm[3]
> + g2 = ccm[4]
> + b1 = ccm[6]
> + b2 = ccm[7]
> + '''
> + COLOR MATRIX LOOKS AS BELOW
> + R1 R2 R3 Rval Outr
> + G1 G2 G3 * Gval = G
> + B1 B2 B3 Bval B
> + Will be optimising 6 elements and working out the third element using 1-r1-r2 = r3
> + '''
> +
> + x0 = [r1, r2, g1, g2, b1, b2]
> + '''
> + We use our old CCM as the initial guess for the program to find the
> + optimised matrix
> + '''
> + result = minimize(guess, x0, args=(r, g, b, m_lab), tol=0.01)
> + '''
> + This produces a color matrix which has the lowest delta E possible,
> + based off the input data. Note it is impossible for this to reach
> + zero since the input data is imperfect
> + '''
> +
> + Cam.log += ("\n \n Optimised Matrix Below: \n \n")
> + [r1, r2, g1, g2, b1, b2] = result.x
> + # The new, optimised color correction matrix values
> + optimised_ccm = [r1, r2, (1 - r1 - r2), g1, g2, (1 - g1 - g2), b1, b2, (1 - b1 - b2)]
> +
> + # This is the optimised Color Matrix (preserving greys by summing rows up to 1)
> + Cam.log += str(optimised_ccm)
> + Cam.log += "\n Old Color Correction Matrix Below \n"
> + Cam.log += str(ccm)
> +
> + formatted_ccm = np.array(original_ccm).reshape((3, 3))
> +
> + '''
> + below is a whole load of code that then applies the latest color
> + matrix, and returns LAB values for color. This can then be used
> + to calculate the final delta E
> + '''
> + optimised_ccm_rgb = [] # Original Color Corrected Matrix RGB / LAB
> + optimised_ccm_lab = []
> +
> + formatted_optimised_ccm = np.array(optimised_ccm).reshape((3, 3))
> + after_gamma_rgb = []
> + after_gamma_lab = []
> +
> + for RGB in zip(r, g, b):
> + ccm_applied_rgb = np.dot(formatted_ccm, (np.array(RGB) / 256))
> + optimised_ccm_rgb.append(gamma(ccm_applied_rgb))
> + optimised_ccm_lab.append(colors.RGB_to_LAB(ccm_applied_rgb))
> +
> + optimised_ccm_applied_rgb = np.dot(formatted_optimised_ccm, np.array(RGB) / 256)
> + after_gamma_rgb.append(gamma(optimised_ccm_applied_rgb))
> + after_gamma_lab.append(colors.RGB_to_LAB(optimised_ccm_applied_rgb))
> + '''
> + Gamma After RGB / LAB - not used in calculations, only used for visualisation
> + We now want to spit out some data that shows
> + how the optimisation has improved the color matrices
> + '''
> + Cam.log += "Here are the Improvements"
> +
> + # CALCULATE WORST CASE delta e
> + old_worst_delta_e = 0
> + before_average = transform_and_evaluate(formatted_ccm, r, g, b, m_lab)
> + new_worst_delta_e = 0
> + after_average = transform_and_evaluate(formatted_optimised_ccm, r, g, b, m_lab)
> + for i in range(24):
> + old_delta_e = deltae(optimised_ccm_lab[i], m_lab[i]) # Current Old Delta E
> + new_delta_e = deltae(after_gamma_lab[i], m_lab[i]) # Current New Delta E
> + if old_delta_e > old_worst_delta_e:
> + old_worst_delta_e = old_delta_e
> + if new_delta_e > new_worst_delta_e:
> + new_worst_delta_e = new_delta_e
> +
> + Cam.log += "Before color correction matrix was optimised, we got an average delta E of " + str(before_average) + " and a maximum delta E of " + str(old_worst_delta_e)
> + Cam.log += "After color correction matrix was optimised, we got an average delta E of " + str(after_average) + " and a maximum delta E of " + str(new_worst_delta_e)
> +
> + visualise_macbeth_chart(m_rgb, optimised_ccm_rgb, after_gamma_rgb, str(Img.col) + str(matrix_selection_types[typenum]))
> + '''
> + The program will also save some visualisations of improvements.
> + Very pretty to look at. Top rectangle is ideal, Left square is
> + before optimisation, right square is after.
> + '''
> +
> + """
> + if a ccm has already been calculated for that temperature then don't
> + overwrite but save both. They will then be averaged later on
> + """ # Now going to use optimised color matrix, optimised_ccm
> + if Img.col in ccm_tab.keys():
> + ccm_tab[Img.col].append(optimised_ccm)
> + else:
> + ccm_tab[Img.col] = [optimised_ccm]
> + Cam.log += '\n'
> +
> + Cam.log += '\nFinished processing images'
> + """
> + average any ccms that share a colour temperature
> + """
> + for k, v in ccm_tab.items():
> + tab = np.mean(v, axis=0)
> + tab = np.where((10000 * tab) % 1 <= 0.05, tab + 0.00001, tab)
> + tab = np.where((10000 * tab) % 1 >= 0.95, tab - 0.00001, tab)
> + ccm_tab[k] = list(np.round(tab, 5))
> + Cam.log += '\nMatrix calculated for colour temperature of {} K'.format(k)
> +
> + """
> + return all ccms with respective colour temperature in the correct format,
> + sorted by their colour temperature
> + """
> + sorted_ccms = sorted(ccm_tab.items(), key=lambda kv: kv[0])
> + ccms = []
> + for i in sorted_ccms:
> + ccms.append({
> + 'ct': i[0],
> + 'ccm': i[1]
> + })
> + return ccms
> +
> +
> +def guess(x0, r, g, b, m_lab): # provides a method of numerical feedback for the optimisation code
> + [r1, r2, g1, g2, b1, b2] = x0
> + ccm = np.array([r1, r2, (1 - r1 - r2),
> + g1, g2, (1 - g1 - g2),
> + b1, b2, (1 - b1 - b2)]).reshape((3, 3)) # format the matrix correctly
> + return transform_and_evaluate(ccm, r, g, b, m_lab)
> +
> +
> +def transform_and_evaluate(ccm, r, g, b, m_lab): # Transforms colors to LAB and applies the correction matrix
> + # create list of matrix changed colors
> + realrgb = []
> + for RGB in zip(r, g, b):
> + rgb_post_ccm = np.dot(ccm, np.array(RGB) / 256) # This is RGB values after the color correction matrix has been applied
> + realrgb.append(colors.RGB_to_LAB(rgb_post_ccm))
> + # now compare that with m_lab and return numeric result, averaged for each patch
> + return (sumde(realrgb, m_lab) / 24) # returns an average result of delta E
> +
> +
> +def sumde(listA, listB):
> + global typenum, test_patches
> + sumde = 0
> + maxde = 0
> + patchde = [] # Create array of the delta E values for each patch. useful for optimisation of certain patches
> + for listA_item, listB_item in zip(listA, listB):
> + if maxde < (deltae(listA_item, listB_item)):
> + maxde = deltae(listA_item, listB_item)
> + patchde.append(deltae(listA_item, listB_item))
> + sumde += deltae(listA_item, listB_item)
> + '''
> + The different options specified at the start allow for
> + the maximum to be returned, average or specific patches
> + '''
> + if typenum == 0:
> + return sumde
> + if typenum == 1:
> + return maxde
> + if typenum == 2:
> + output = sum([patchde[test_patch] for test_patch in test_patches])
> + # Selects only certain patches and returns the output for them
> + return output
> +
> +
> +"""
> +calculates the ccm for an individual image.
> +ccms are calculated in rgb space, and are fit by hand. Although it is a 3x3
> +matrix, each row must add up to 1 in order to conserve greyness, simplifying
> +calculation.
> +The initial CCM is calculated in RGB, and then optimised in LAB color space
> +This simplifies the initial calculation but then gets us the accuracy of
> +using LAB color space.
> +"""
> +
> +
> +def do_ccm(r, g, b, m_srgb):
> + rb = r-b
> + gb = g-b
> + rb_2s = (rb * rb)
> + rb_gbs = (rb * gb)
> + gb_2s = (gb * gb)
> +
> + r_rbs = rb * (m_srgb[..., 0] - b)
> + r_gbs = gb * (m_srgb[..., 0] - b)
> + g_rbs = rb * (m_srgb[..., 1] - b)
> + g_gbs = gb * (m_srgb[..., 1] - b)
> + b_rbs = rb * (m_srgb[..., 2] - b)
> + b_gbs = gb * (m_srgb[..., 2] - b)
> +
> + """
> + Obtain least squares fit
> + """
> + rb_2 = np.sum(rb_2s)
> + gb_2 = np.sum(gb_2s)
> + rb_gb = np.sum(rb_gbs)
> + r_rb = np.sum(r_rbs)
> + r_gb = np.sum(r_gbs)
> + g_rb = np.sum(g_rbs)
> + g_gb = np.sum(g_gbs)
> + b_rb = np.sum(b_rbs)
> + b_gb = np.sum(b_gbs)
> +
> + det = rb_2 * gb_2 - rb_gb * rb_gb
> +
> + """
> + Raise error if matrix is singular...
> + This shouldn't really happen with real data but if it does just take new
> + pictures and try again, not much else to be done unfortunately...
> + """
> + if det < 0.001:
> + raise ArithmeticError
> +
> + r_a = (gb_2 * r_rb - rb_gb * r_gb) / det
> + r_b = (rb_2 * r_gb - rb_gb * r_rb) / det
> + """
> + Last row can be calculated by knowing the sum must be 1
> + """
> + r_c = 1 - r_a - r_b
> +
> + g_a = (gb_2 * g_rb - rb_gb * g_gb) / det
> + g_b = (rb_2 * g_gb - rb_gb * g_rb) / det
> + g_c = 1 - g_a - g_b
> +
> + b_a = (gb_2 * b_rb - rb_gb * b_gb) / det
> + b_b = (rb_2 * b_gb - rb_gb * b_rb) / det
> + b_c = 1 - b_a - b_b
> +
> + """
> + format ccm
> + """
> + ccm = [r_a, r_b, r_c, g_a, g_b, g_c, b_a, b_b, b_c]
> +
> + return ccm
> +
> +
> +def deltae(colorA, colorB):
> + return ((colorA[0] - colorB[0]) ** 2 + (colorA[1] - colorB[1]) ** 2 + (colorA[2] - colorB[2]) ** 2) ** 0.5
> + # return ((colorA[1]-colorB[1]) * * 2 + (colorA[2]-colorB[2]) * * 2) * * 0.5
> + # UNCOMMENT IF YOU WANT TO NEGLECT LUMINANCE FROM CALCULATION OF DELTA E
> diff --git a/utils/tuning/libtuning/ctt_colors.py b/utils/tuning/libtuning/ctt_colors.py
> new file mode 100644
> index 000000000000..cb4d236b04d7
> --- /dev/null
> +++ b/utils/tuning/libtuning/ctt_colors.py
> @@ -0,0 +1,30 @@
> +# Program to convert from RGB to LAB color space
> +def RGB_to_LAB(RGB): # where RGB is a 1x3 array. e.g RGB = [100, 255, 230]
> + num = 0
> + XYZ = [0, 0, 0]
> + # converted all the three R, G, B to X, Y, Z
> + X = RGB[0] * 0.4124 + RGB[1] * 0.3576 + RGB[2] * 0.1805
> + Y = RGB[0] * 0.2126 + RGB[1] * 0.7152 + RGB[2] * 0.0722
> + Z = RGB[0] * 0.0193 + RGB[1] * 0.1192 + RGB[2] * 0.9505
> +
> + XYZ[0] = X / 255 * 100
> + XYZ[1] = Y / 255 * 100 # XYZ Must be in range 0 -> 100, so scale down from 255
> + XYZ[2] = Z / 255 * 100
> + XYZ[0] = XYZ[0] / 95.047 # ref_X = 95.047 Observer= 2°, Illuminant= D65
> + XYZ[1] = XYZ[1] / 100.0 # ref_Y = 100.000
> + XYZ[2] = XYZ[2] / 108.883 # ref_Z = 108.883
> + num = 0
> + for value in XYZ:
> + if value > 0.008856:
> + value = value ** (0.3333333333333333)
> + else:
> + value = (7.787 * value) + (16 / 116)
> + XYZ[num] = value
> + num = num + 1
> +
> + # L, A, B, values calculated below
> + L = (116 * XYZ[1]) - 16
> + a = 500 * (XYZ[0] - XYZ[1])
> + b = 200 * (XYZ[1] - XYZ[2])
> +
> + return [L, a, b]
> diff --git a/utils/tuning/libtuning/ctt_ransac.py b/utils/tuning/libtuning/ctt_ransac.py
> new file mode 100644
> index 000000000000..01bba3022ef0
> --- /dev/null
> +++ b/utils/tuning/libtuning/ctt_ransac.py
> @@ -0,0 +1,71 @@
> +# SPDX-License-Identifier: BSD-2-Clause
> +#
> +# Copyright (C) 2019, Raspberry Pi Ltd
> +#
> +# camera tuning tool RANSAC selector for Macbeth chart locator
> +
> +import numpy as np
> +
> +scale = 2
> +
> +
> +"""
> +constructs normalised macbeth chart corners for ransac algorithm
> +"""
> +def get_square_verts(c_err=0.05, scale=scale):
> + """
> + define macbeth chart corners
> + """
> + b_bord_x, b_bord_y = scale*8.5, scale*13
> + s_bord = 6*scale
> + side = 41*scale
> + x_max = side*6 + 5*s_bord + 2*b_bord_x
> + y_max = side*4 + 3*s_bord + 2*b_bord_y
> + c1 = (0, 0)
> + c2 = (0, y_max)
> + c3 = (x_max, y_max)
> + c4 = (x_max, 0)
> + mac_norm = np.array((c1, c2, c3, c4), np.float32)
> + mac_norm = np.array([mac_norm])
> +
> + square_verts = []
> + square_0 = np.array(((0, 0), (0, side),
> + (side, side), (side, 0)), np.float32)
> + offset_0 = np.array((b_bord_x, b_bord_y), np.float32)
> + c_off = side * c_err
> + offset_cont = np.array(((c_off, c_off), (c_off, -c_off),
> + (-c_off, -c_off), (-c_off, c_off)), np.float32)
> + square_0 += offset_0
> + square_0 += offset_cont
> + """
> + define macbeth square corners
> + """
> + for i in range(6):
> + shift_i = np.array(((i*side, 0), (i*side, 0),
> + (i*side, 0), (i*side, 0)), np.float32)
> + shift_bord = np.array(((i*s_bord, 0), (i*s_bord, 0),
> + (i*s_bord, 0), (i*s_bord, 0)), np.float32)
> + square_i = square_0 + shift_i + shift_bord
> + for j in range(4):
> + shift_j = np.array(((0, j*side), (0, j*side),
> + (0, j*side), (0, j*side)), np.float32)
> + shift_bord = np.array(((0, j*s_bord),
> + (0, j*s_bord), (0, j*s_bord),
> + (0, j*s_bord)), np.float32)
> + square_j = square_i + shift_j + shift_bord
> + square_verts.append(square_j)
> + # print('square_verts')
> + # print(square_verts)
> + return np.array(square_verts, np.float32), mac_norm
> +
> +
> +def get_square_centres(c_err=0.05, scale=scale):
> + """
> + define macbeth square centres
> + """
> + verts, mac_norm = get_square_verts(c_err, scale=scale)
> +
> + centres = np.mean(verts, axis=1)
> + # print('centres')
> + # print(centres)
> + return np.array(centres, np.float32)
--
Regards,
Laurent Pinchart
More information about the libcamera-devel
mailing list