libcamera/pinchartl/libcamera.git - Laurent Pinchart's clone of libcamera

path: root/src/apps/qcam/cam_select_dialog.cpp

diff options

Diffstat (limited to 'src/apps/qcam/cam_select_dialog.cpp')

0 files changed, 0 insertions, 0 deletions

44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181

# SPDX-License-Identifier: BSD-2-Clause
#
# Copyright (C) 2019, Raspberry Pi (Trading) Limited
#
# ctt_geq.py - camera tuning tool for GEQ (green equalisation)

from ctt_tools import *
import matplotlib.pyplot as plt
import scipy.optimize as optimize


"""
Uses green differences in macbeth patches to fit green equalisation threshold
model. Ideally, all macbeth chart centres would fall below the threshold as
these should be corrected by geq.
"""
def geq_fit(Cam, plot):
    imgs = Cam.imgs
    """
    green equalisation to mitigate mazing.
    Fits geq model by looking at difference
    between greens in macbeth patches
    """
    geqs = np.array([geq(Cam, Img)*Img.againQ8_norm for Img in imgs])
    Cam.log += '\nProcessed all images'
    geqs = geqs.reshape((-1, 2))
    """
    data is sorted by green difference and top half is selected since higher
    green difference data define the decision boundary.
    """
    geqs = np.array(sorted(geqs, key=lambda r: np.abs((r[1]-r[0])/r[0])))

    length = len(geqs)
    g0 = geqs[length//2:, 0]
    g1 = geqs[length//2:, 1]
    gdiff = np.abs(g0-g1)
    """
    find linear fit by minimising asymmetric least square errors
    in order to cover most of the macbeth images.
    the philosophy here is that every macbeth patch should fall within the
    threshold, hence the upper bound approach
    """
    def f(params):
        m, c = params
        a = gdiff - (m*g0+c)
        """
        asymmetric square error returns:
            1.95 * a**2 if a is positive
            0.05 * a**2 if a is negative
        """
        return(np.sum(a**2+0.95*np.abs(a)*a))

    initial_guess = [0.01, 500]
    """
    Nelder-Mead is usually not the most desirable optimisation method
    but has been chosen here due to its robustness to undifferentiability
    (is that a word?)
    """
    result = optimize.minimize(f, initial_guess, method='Nelder-Mead')
    """
    need to check if the fit worked correectly
    """
    if result.success:
        slope, offset = result.x
        Cam.log += '\nFit result: slope = {:.5f} '.format(slope)
        Cam.log += 'offset = {}'.format(int(offset))
        """
        optional plotting code
        """
        if plot:
            x = np.linspace(max(g0)*1.1, 100)
            y = slope*x + offset
            plt.title('GEQ Asymmetric \'Upper Bound\' Fit')
            plt.plot(x, y, color='red', ls='--', label='fit')
            plt.scatter(g0, gdiff, color='b', label='data')
            plt.ylabel('Difference in green channels')
            plt.xlabel('Green value')

        """
        This upper bound asymmetric gives correct order of magnitude values.
        The pipeline approximates a 1st derivative of a gaussian with some
        linear piecewise functions, introducing arbitrary cutoffs. For
        pessimistic geq, the model parameters have been increased by a
        scaling factor/constant.

        Feel free to tune these or edit the json files directly if you
        belive there are still mazing effects left (threshold too low) or if you
        think it is being overcorrected (threshold too high).
        We have gone for a one size fits most approach that will produce
        acceptable results in most applications.
        """
        slope *= 1.5
        offset += 201
        Cam.log += '\nFit after correction factors: slope = {:.5f}'.format(slope)
        Cam.log += ' offset = {}'.format(int(offset))
        """
        clamp offset at 0 due to pipeline considerations
        """
        if offset < 0:
            Cam.log += '\nOffset raised to 0'
            offset = 0
        """
        optional plotting code
        """
        if plot:
            y2 = slope*x + offset
            plt.plot(x, y2, color='green', ls='--', label='scaled fit')
            plt.grid()
            plt.legend()
            plt.show()

        """
    the case where for some reason the fit didn't work correctly

    Transpose data and then least squares linear fit. Transposing data
    makes it robust to many patches where green difference is the same
    since they only contribute to one error minimisation, instead of dragging
    the entire linear fit down.
    """

    else:
        print('\nError! Couldn\'t fit asymmetric lest squares')
        print(result.message)
        Cam.log += '\nWARNING: Asymmetric least squares fit failed! '
        Cam.log += 'Standard fit used could possibly lead to worse results'
        fit = np.polyfit(gdiff, g0, 1)
        offset, slope = -fit[1]/fit[0], 1/fit[0]
        Cam.log += '\nFit result: slope = {:.5f} '.format(slope)
        Cam.log += 'offset = {}'.format(int(offset))
        """
        optional plotting code
        """
        if plot:
            x = np.linspace(max(g0)*1.1, 100)
            y = slope*x + offset
            plt.title('GEQ Linear Fit')
            plt.plot(x, y, color='red', ls='--', label='fit')
            plt.scatter(g0, gdiff, color='b', label='data')
            plt.ylabel('Difference in green channels')
            plt.xlabel('Green value')
        """
        Scaling factors (see previous justification)
        The model here will not be an upper bound so scaling factors have
        been increased.
        This method of deriving geq model parameters is extremely arbitrary
        and undesirable.
        """
        slope *= 2.5
        offset += 301


context:
space:
mode: