diff options
Diffstat (limited to 'utils/tuning/libtuning/ctt_awb.py')
| -rw-r--r-- | utils/tuning/libtuning/ctt_awb.py | 378 |
1 files changed, 378 insertions, 0 deletions
diff --git a/utils/tuning/libtuning/ctt_awb.py b/utils/tuning/libtuning/ctt_awb.py new file mode 100644 index 00000000..abf22321 --- /dev/null +++ b/utils/tuning/libtuning/ctt_awb.py @@ -0,0 +1,378 @@ +# SPDX-License-Identifier: BSD-2-Clause +# +# Copyright (C) 2019, Raspberry Pi Ltd +# +# camera tuning tool for AWB + +import matplotlib.pyplot as plt +from bisect import bisect_left +from scipy.optimize import fmin +import numpy as np + +from .image import Image + + +""" +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.color + 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 |