/* SPDX-License-Identifier: BSD-2-Clause */ /* * Copyright (C) 2019, Raspberry Pi (Trading) Limited * * agc.cpp - AGC/AEC control algorithm */ #include #include "linux/bcm2835-isp.h" #include "../awb_status.h" #include "../device_status.h" #include "../histogram.hpp" #include "../logging.hpp" #include "../lux_status.h" #include "../metadata.hpp" #include "agc.hpp" using namespace RPi; #define NAME "rpi.agc" #define PIPELINE_BITS 13 // seems to be a 13-bit pipeline void AgcMeteringMode::Read(boost::property_tree::ptree const ¶ms) { int num = 0; for (auto &p : params.get_child("weights")) { if (num == AGC_STATS_SIZE) throw std::runtime_error("AgcConfig: too many weights"); weights[num++] = p.second.get_value(); } if (num != AGC_STATS_SIZE) throw std::runtime_error("AgcConfig: insufficient weights"); } static std::string read_metering_modes(std::map &metering_modes, boost::property_tree::ptree const ¶ms) { std::string first; for (auto &p : params) { AgcMeteringMode metering_mode; metering_mode.Read(p.second); metering_modes[p.first] = std::move(metering_mode); if (first.empty()) first = p.first; } return first; } static int read_double_list(std::vector &list, boost::property_tree::ptree const ¶ms) { for (auto &p : params) list.push_back(p.second.get_value()); return list.size(); } void AgcExposureMode::Read(boost::property_tree::ptree const ¶ms) { int num_shutters = read_double_list(shutter, params.get_child("shutter")); int num_ags = read_double_list(gain, params.get_child("gain")); if (num_shutters < 2 || num_ags < 2) throw std::runtime_error( "AgcConfig: must have at least two entries in exposure profile"); if (num_shutters != num_ags) throw std::runtime_error( "AgcConfig: expect same number of exposure and gain entries in exposure profile"); } static std::string read_exposure_modes(std::map &exposure_modes, boost::property_tree::ptree const ¶ms) { std::string first; for (auto &p : params) { AgcExposureMode exposure_mode; exposure_mode.Read(p.second); exposure_modes[p.first] = std::move(exposure_mode); if (first.empty()) first = p.first; } return first; } void AgcConstraint::Read(boost::property_tree::ptree const ¶ms) { std::string bound_string = params.get("bound", ""); transform(bound_string.begin(), bound_string.end(), bound_string.begin(), ::toupper); if (bound_string != "UPPER" && bound_string != "LOWER") throw std::runtime_error( "AGC constraint type should be UPPER or LOWER"); bound = bound_string == "UPPER" ? Bound::UPPER : Bound::LOWER; q_lo = params.get("q_lo"); q_hi = params.get("q_hi"); Y_target.Read(params.get_child("y_target")); } static AgcConstraintMode read_constraint_mode(boost::property_tree::ptree const ¶ms) { AgcConstraintMode mode; for (auto &p : params) { AgcConstraint constraint; constraint.Read(p.second); mode.push_back(std::move(constraint)); } return mode; } static std::string read_constraint_modes( std::map &constraint_modes, boost::property_tree::ptree const ¶ms) { std::string first; for (auto &p : params) { constraint_modes[p.first] = read_constraint_mode(p.second); if (first.empty()) first = p.first; } return first; } void AgcConfig::Read(boost::property_tree::ptree const ¶ms) { RPI_LOG("AgcConfig"); default_metering_mode = read_metering_modes( metering_modes, params.get_child("metering_modes")); default_exposure_mode = read_exposure_modes( exposure_modes, params.get_child("exposure_modes")); default_constraint_mode = read_constraint_modes( constraint_modes, params.get_child("constraint_modes")); Y_target.Read(params.get_child("y_target")); speed = params.get("speed", 0.2); startup_frames = params.get("startup_frames", 10); fast_reduce_threshold = params.get("fast_reduce_threshold", 0.4); base_ev = params.get("base_ev", 1.0); } Agc::Agc(Controller *controller) : AgcAlgorithm(controller), metering_mode_(nullptr), exposure_mode_(nullptr), constraint_mode_(nullptr), frame_count_(0), lock_count_(0) { ev_ = status_.ev = 1.0; flicker_period_ = status_.flicker_period = 0.0; fixed_shutter_ = status_.fixed_shutter = 0; fixed_analogue_gain_ = status_.fixed_analogue_gain = 0.0; // set to zero initially, so we can tell it's not been calculated status_.total_exposure_value = 0.0; status_.target_exposure_value = 0.0; status_.locked = false; output_status_ = status_; } char const *Agc::Name() const { return NAME; } void Agc::Read(boost::property_tree::ptree const ¶ms) { RPI_LOG("Agc"); config_.Read(params); // Set the config's defaults (which are the first ones it read) as our // current modes, until someone changes them. (they're all known to // exist at this point) metering_mode_name_ = config_.default_metering_mode; metering_mode_ = &config_.metering_modes[metering_mode_name_]; exposure_mode_name_ = config_.default_exposure_mode; exposure_mode_ = &config_.exposure_modes[exposure_mode_name_]; constraint_mode_name_ = config_.default_constraint_mode; constraint_mode_ = &config_.constraint_modes[constraint_mode_name_]; } void Agc::SetEv(double ev) { std::unique_lock lock(settings_mutex_); ev_ = ev; } void Agc::SetFlickerPeriod(double flicker_period) { std::unique_lock lock(settings_mutex_); flicker_period_ = flicker_period; } void Agc::SetFixedShutter(double fixed_shutter) { std::unique_lock lock(settings_mutex_); fixed_shutter_ = fixed_shutter; } void Agc::SetFixedAnalogueGain(double fixed_analogue_gain) { std::unique_lock lock(settings_mutex_); fixed_analogue_gain_ = fixed_analogue_gain; } void Agc::SetMeteringMode(std::string const &metering_mode_name) { std::unique_lock lock(settings_mutex_); metering_mode_name_ = metering_mode_name; } void Agc::SetExposureMode(std::string const &exposure_mode_name) { std::unique_lock lock(settings_mutex_); exposure_mode_name_ = exposure_mode_name; } void Agc::SetConstraintMode(std::string const &constraint_mode_name) { std::unique_lock lock(settings_mutex_); constraint_mode_name_ = constraint_mode_name; } void Agc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) { // On a mode switch, it's possible the exposure profile could change, // so we run through the dividing up of exposure/gain again and // write the results into the metadata we've been given. if (status_.total_exposure_value) { housekeepConfig(); divvyupExposure(); writeAndFinish(metadata, false); } } void Agc::Prepare(Metadata *image_metadata) { AgcStatus status; { std::unique_lock lock(output_mutex_); status = output_status_; } int lock_count = lock_count_; lock_count_ = 0; status.digital_gain = 1.0; if (status_.total_exposure_value) { // Process has run, so we have meaningful values. DeviceStatus device_status; if (image_metadata->Get("device.status", device_status) == 0) { double actual_exposure = device_status.shutter_speed * device_status.analogue_gain; if (actual_exposure) { status.digital_gain = status_.total_exposure_value / actual_exposure; RPI_LOG("Want total exposure " << status_.total_exposure_value); // Never ask for a gain < 1.0, and also impose // some upper limit. Make it customisable? status.digital_gain = std::max( 1.0, std::min(status.digital_gain, 4.0)); RPI_LOG("Actual exposure " << actual_exposure); RPI_LOG("Use digital_gain " << status.digital_gain); RPI_LOG("Effective exposure " << actual_exposure * status.digital_gain); // Decide whether AEC/AGC has converged. // Insist AGC is steady for MAX_LOCK_COUNT // frames before we say we are "locked". // (The hard-coded constants may need to // become customisable.) if (status.target_exposure_value) { #define MAX_LOCK_COUNT 3 double err = 0.10 * status.target_exposure_value + 200; if (actual_exposure < status.target_exposure_value + err && actual_exposure > status.target_exposure_value - err) lock_count_ = std::min(lock_count + 1, MAX_LOCK_COUNT); else if (actual_exposure < status.target_exposure_value + 1.5 * err && actual_exposure > status.target_exposure_value - 1.5 * err) lock_count_ = lock_count; RPI_LOG("Lock count: " << lock_count_); } } } else RPI_LOG(Name() << ": no device metadata"); status.locked = lock_count_ >= MAX_LOCK_COUNT; //printf("%s\n", status.locked ? "+++++++++" : "-"); image_metadata->Set("agc.status", status); } } void Agc::Process(StatisticsPtr &stats, Metadata *image_metadata) { frame_count_++; // First a little bit of housekeeping, fetching up-to-date settings and // configuration, that kind of thing. housekeepConfig(); // Get the current exposure values for the frame that's just arrived. fetchCurrentExposure(image_metadata); // Compute the total gain we require relative to the current exposure. double gain, target_Y; computeGain(stats.get(), image_metadata, gain, target_Y); // Now compute the target (final) exposure which we think we want. computeTargetExposure(gain); // Some of the exposure has to be applied as digital gain, so work out // what that is. This function also tells us whether it's decided to // "desaturate" the image more quickly. bool desaturate = applyDigitalGain(image_metadata, gain, target_Y); // The results have to be filtered so as not to change too rapidly. filterExposure(desaturate); // The last thing is to divvy up the exposure value into a shutter time // and analogue_gain, according to the current exposure mode. divvyupExposure(); // Finally advertise what we've done. writeAndFinish(image_metadata, desaturate); } static void copy_string(std::string const &s, char *d, size_t size) { size_t length = s.copy(d, size - 1); d[length] = '\0'; } void Agc::housekeepConfig() { // First fetch all the up-to-date settings, so no one else has to do it. std::string new_exposure_mode_name, new_constraint_mode_name, new_metering_mode_name; { std::unique_lock lock(settings_mutex_); new_metering_mode_name = metering_mode_name_; new_exposure_mode_name = exposure_mode_name_; new_constraint_mode_name = constraint_mode_name_; status_.ev = ev_; status_.fixed_shutter = fixed_shutter_; status_.fixed_analogue_gain = fixed_analogue_gain_; status_.flicker_period = flicker_period_; } RPI_LOG("ev " << status_.ev << " fixed_shutter " << status_.fixed_shutter << " fixed_analogue_gain " << status_.fixed_analogue_gain); // Make sure the "mode" pointers point to the up-to-date things, if // they've changed. if (strcmp(new_metering_mode_name.c_str(), status_.metering_mode)) { auto it = config_.metering_modes.find(new_metering_mode_name); if (it == config_.metering_modes.end()) throw std::runtime_error("Agc: no metering mode " + new_metering_mode_name); metering_mode_ = &it->second; copy_string(new_metering_mode_name, status_.metering_mode, sizeof(status_.metering_mode)); } if (strcmp(new_exposure_mode_name.c_str(), status_.exposure_mode)) { auto it = config_.exposure_modes.find(new_exposure_mode_name); if (it == config_.exposure_modes.end()) throw std::runtime_error("Agc: no exposure profile " + new_exposure_mode_name); exposure_mode_ = &it->second; copy_string(new_exposure_mode_name, status_.exposure_mode, sizeof(status_.exposure_mode)); } if (strcmp(new_constraint_mode_name.c_str(), status_.constraint_mode)) { auto it = config_.constraint_modes.find(new_constraint_mode_name); if (it == config_.constraint_modes.end()) throw std::runtime_error("Agc: no constraint list " + new_constraint_mode_name); constraint_mode_ = &it->second; copy_string(new_constraint_mode_name, status_.constraint_mode, sizeof(status_.constraint_mode)); } RPI_LOG("exposure_mode " << new_exposure_mode_name << " constraint_mode " << new_constraint_mode_name << " metering_mode " << new_metering_mode_name); } void Agc::fetchCurrentExposure(Metadata *image_metadata) { std::unique_lock lock(*image_metadata); DeviceStatus *device_status = image_metadata->GetLocked("device.status"); if (!device_status) throw std::runtime_error("Agc: no device metadata"); current_.shutter = device_status->shutter_speed; current_.analogue_gain = device_status->analogue_gain; AgcStatus *agc_status = image_metadata->GetLocked("agc.status"); current_.total_exposure = agc_status ? agc_status->total_exposure_value : 0; current_.total_exposure_no_dg = current_.shutter * current_.analogue_gain; } static double compute_initial_Y(bcm2835_isp_stats *stats, Metadata *image_metadata, double weights[]) { bcm2835_isp_stats_region *regions = stats->agc_stats; struct AwbStatus awb; awb.gain_r = awb.gain_g = awb.gain_b = 1.0; // in case no metadata if (image_metadata->Get("awb.status", awb) != 0) RPI_WARN("Agc: no AWB status found"); double Y_sum = 0, weight_sum = 0; for (int i = 0; i < AGC_STATS_SIZE; i++) { if (regions[i].counted == 0) continue; weight_sum += weights[i]; double Y = regions[i].r_sum * awb.gain_r * .299 + regions[i].g_sum * awb.gain_g * .587 + regions[i].b_sum * awb.gain_b * .114; Y /= regions[i].counted; Y_sum += Y * weights[i]; } return Y_sum / weight_sum / (1 << PIPELINE_BITS); } // We handle extra gain through EV by adjusting our Y targets. However, you // simply can't monitor histograms once they get very close to (or beyond!) // saturation, so we clamp the Y targets to this value. It does mean that EV // increases don't necessarily do quite what you might expect in certain // (contrived) cases. #define EV_GAIN_Y_TARGET_LIMIT 0.9 static double constraint_compute_gain(AgcConstraint &c, Histogram &h, double lux, double ev_gain, double &target_Y) { target_Y = c.Y_target.Eval(c.Y_target.Domain().Clip(lux)); target_Y = std::min(EV_GAIN_Y_TARGET_LIMIT, target_Y * ev_gain); double iqm = h.InterQuantileMean(c.q_lo, c.q_hi); return (target_Y * NUM_HISTOGRAM_BINS) / iqm; } void Agc::computeGain(bcm2835_isp_stats *statistics, Metadata *image_metadata, double &gain, double &target_Y) { struct LuxStatus lux = {}; lux.lux = 400; // default lux level to 400 in case no metadata found if (image_metadata->Get("lux.status", lux) != 0) RPI_WARN("Agc: no lux level found"); Histogram h(statistics->hist[0].g_hist, NUM_HISTOGRAM_BINS); double ev_gain = status_.ev * config_.base_ev; // The initial gain and target_Y come from some of the regions. After // that we consider the histogram constraints. target_Y = config_.Y_target.Eval(config_.Y_target.Domain().Clip(lux.lux)); target_Y = std::min(EV_GAIN_Y_TARGET_LIMIT, target_Y * ev_gain); double initial_Y = compute_initial_Y(statistics, image_metadata, metering_mode_->weights); gain = std::min(10.0, target_Y / (initial_Y + .001)); RPI_LOG("Initially Y " << initial_Y << " target " << target_Y << " gives gain " << gain); for (auto &c : *constraint_mode_) { double new_target_Y; double new_gain = constraint_compute_gain(c, h, lux.lux, ev_gain, new_target_Y); RPI_LOG("Constraint has target_Y " << new_target_Y << " giving gain " << new_gain); if (c.bound == AgcConstraint::Bound::LOWER && new_gain > gain) { RPI_LOG("Lower bound constraint adopted"); gain = new_gain, target_Y = new_target_Y; } else if (c.bound == AgcConstraint::Bound::UPPER && new_gain < gain) { RPI_LOG("Upper bound constraint adopted"); gain = new_gain, target_Y = new_target_Y; } } RPI_LOG("Final gain " << gain << " (target_Y " << target_Y << " ev " << status_.ev << " base_ev " << config_.base_ev << ")"); } void Agc::computeTargetExposure(double gain) { // The statistics reflect the image without digital gain, so the final // total exposure we're aiming for is: target_.total_exposure = current_.total_exposure_no_dg * gain; // The final target exposure is also limited to what the exposure // mode allows. double max_total_exposure = (status_.fixed_shutter != 0.0 ? status_.fixed_shutter : exposure_mode_->shutter.back()) * (status_.fixed_analogue_gain != 0.0 ? status_.fixed_analogue_gain : exposure_mode_->gain.back()); target_.total_exposure = std::min(target_.total_exposure, max_total_exposure); RPI_LOG("Target total_exposure " << target_.total_exposure); } bool Agc::applyDigitalGain(Metadata *image_metadata, double gain, double target_Y) { double dg = 1.0; // I think this pipeline subtracts black level and rescales before we // get the stats, so no need to worry about it. struct AwbStatus awb; if (image_metadata->Get("awb.status", awb) == 0) { double min_gain = std::min(awb.gain_r, std::min(awb.gain_g, awb.gain_b)); dg *= std::max(1.0, 1.0 / min_gain); } else RPI_WARN("Agc: no AWB status found"); RPI_LOG("after AWB, target dg " << dg << " gain " << gain << " target_Y " << target_Y); // Finally, if we're trying to reduce exposure but the target_Y is // "close" to 1.0, then the gain computed for that constraint will be // only slightly less than one, because the measured Y can never be // larger than 1.0. When this happens, demand a large digital gain so // that the exposure can be reduced, de-saturating the image much more // quickly (and we then approach the correct value more quickly from // below). bool desaturate = target_Y > config_.fast_reduce_threshold && gain < sqrt(target_Y); if (desaturate) dg /= config_.fast_reduce_threshold; RPI_LOG("Digital gain " << dg << " desaturate? " << desaturate); target_.total_exposure_no_dg = target_.total_exposure / dg; RPI_LOG("Target total_exposure_no_dg " << target_.total_exposure_no_dg); return desaturate; } void Agc::filterExposure(bool desaturate) { double speed = frame_count_ <= config_.startup_frames ? 1.0 : config_.speed; if (filtered_.total_exposure == 0.0) { filtered_.total_exposure = target_.total_exposure; filtered_.total_exposure_no_dg = target_.total_exposure_no_dg; } else { // If close to the result go faster, to save making so many // micro-adjustments on the way. (Make this customisable?) if (filtered_.total_exposure < 1.2 * target_.total_exposure && filtered_.total_exposure > 0.8 * target_.total_exposure) speed = sqrt(speed); filtered_.total_exposure = speed * target_.total_exposure + filtered_.total_exposure * (1.0 - speed); // When desaturing, take a big jump down in exposure_no_dg, // which we'll hide with digital gain. if (desaturate) filtered_.total_exposure_no_dg = target_.total_exposure_no_dg; else filtered_.total_exposure_no_dg = speed * target_.total_exposure_no_dg + filtered_.total_exposure_no_dg * (1.0 - speed); } // We can't let the no_dg exposure deviate too far below the // total exposure, as there might not be enough digital gain available // in the ISP to hide it (which will cause nasty oscillation). if (filtered_.total_exposure_no_dg < filtered_.total_exposure * config_.fast_reduce_threshold) filtered_.total_exposure_no_dg = filtered_.total_exposure * config_.fast_reduce_threshold; RPI_LOG("After filtering, total_exposure " << filtered_.total_exposure << " no dg " << filtered_.total_exposure_no_dg); } void Agc::divvyupExposure() { // Sending the fixed shutter/gain cases through the same code may seem // unnecessary, but it will make more sense when extend this to cover // variable aperture. double exposure_value = filtered_.total_exposure_no_dg; double shutter_time, analogue_gain; shutter_time = status_.fixed_shutter != 0.0 ? status_.fixed_shutter : exposure_mode_->shutter[0]; analogue_gain = status_.fixed_analogue_gain != 0.0 ? status_.fixed_analogue_gain : exposure_mode_->gain[0]; if (shutter_time * analogue_gain < exposure_value) { for (unsigned int stage = 1; stage < exposure_mode_->gain.size(); stage++) { if (status_.fixed_shutter == 0.0) { if (exposure_mode_->shutter[stage] * analogue_gain >= exposure_value) { shutter_time = exposure_value / analogue_gain; break; } shutter_time = exposure_mode_->shutter[stage]; } if (status_.fixed_analogue_gain == 0.0) { if (exposure_mode_->gain[stage] * shutter_time >= exposure_value) { analogue_gain = exposure_value / shutter_time; break; } analogue_gain = exposure_mode_->gain[stage]; } } } RPI_LOG("Divided up shutter and gain are " << shutter_time << " and " << analogue_gain); // Finally adjust shutter time for flicker avoidance (require both // shutter and gain not to be fixed). if (status_.fixed_shutter == 0.0 && status_.fixed_analogue_gain == 0.0 && status_.flicker_period != 0.0) { int flicker_periods = shutter_time / status_.flicker_period; if (flicker_periods > 0) { double new_shutter_time = flicker_periods * status_.flicker_period; analogue_gain *= shutter_time / new_shutter_time; // We should still not allow the ag to go over the // largest value in the exposure mode. Note that this // may force more of the total exposure into the digital // gain as a side-effect. analogue_gain = std::min(analogue_gain, exposure_mode_->gain.back()); shutter_time = new_shutter_time; } RPI_LOG("After flicker avoidance, shutter " << shutter_time << " gain " << analogue_gain); } filtered_.shutter = shutter_time; filtered_.analogue_gain = analogue_gain; } void Agc::writeAndFinish(Metadata *image_metadata, bool desaturate) { status_.total_exposure_value = filtered_.total_exposure; status_.target_exposure_value = desaturate ? 0 : target_.total_exposure_no_dg; status_.shutter_time = filtered_.shutter; status_.analogue_gain = filtered_.analogue_gain; { std::unique_lock lock(output_mutex_); output_status_ = status_; } // Write to metadata as well, in case anyone wants to update the camera // immediately. image_metadata->Set("agc.status", status_); RPI_LOG("Output written, total exposure requested is " << filtered_.total_exposure); RPI_LOG("Camera exposure update: shutter time " << filtered_.shutter << " analogue gain " << filtered_.analogue_gain); } // Register algorithm with the system. static Algorithm *Create(Controller *controller) { return (Algorithm *)new Agc(controller); } static RegisterAlgorithm reg(NAME, &Create); >580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757

# SPDX-License-Identifier: BSD-2-Clause
#
# Copyright (C) 2019, Raspberry Pi Ltd
#
# camera tuning tool Macbeth chart locator

from ctt_ransac import *
from ctt_tools import *
import warnings

"""
NOTE: some custom functions have been used here to make the code more readable.
These are defined in tools.py if they are needed for reference.
"""


"""
Some inconsistencies between packages cause runtime warnings when running
the clustering algorithm. This catches these warnings so they don't flood the
output to the console
"""
def fxn():
    warnings.warn("runtime", RuntimeWarning)


"""
Define the success message
"""
success_msg = 'Macbeth chart located successfully'

def find_macbeth(Cam, img, mac_config=(0, 0)):
    small_chart, show = mac_config
    print('Locating macbeth chart')
    Cam.log += '\nLocating macbeth chart'
    """
    catch the warnings
    """
    warnings.simplefilter("ignore")
    fxn()

    """
    Reference macbeth chart is created that will be correlated with the located
    macbeth chart guess to produce a confidence value for the match.
    """
    ref = cv2.imread(Cam.path + 'ctt_ref.pgm', flags=cv2.IMREAD_GRAYSCALE)
    ref_w = 120
    ref_h = 80
    rc1 = (0, 0)
    rc2 = (0, ref_h)
    rc3 = (ref_w, ref_h)
    rc4 = (ref_w, 0)
    ref_corns = np.array((rc1, rc2, rc3, rc4), np.float32)
    ref_data = (ref, ref_w, ref_h, ref_corns)

    """
    locate macbeth chart
    """
    cor, mac, coords, msg = get_macbeth_chart(img, ref_data)

    # Keep a list that will include this and any brightened up versions of
    # the image for reuse.
    all_images = [img]

    """
    following bits of code tries to fix common problems with simple
    techniques.
    If now or at any point the best correlation is of above 0.75, then
    nothing more is tried as this is a high enough confidence to ensure
    reliable macbeth square centre placement.
    """

    """
    brighten image 2x
    """
    if cor < 0.75:
        a = 2
        img_br = cv2.convertScaleAbs(img, alpha=a, beta=0)
        all_images.append(img_br)
        cor_b, mac_b, coords_b, msg_b = get_macbeth_chart(img_br, ref_data)
        if cor_b > cor:
            cor, mac, coords, msg = cor_b, mac_b, coords_b, msg_b

    """
    brighten image 4x
    """
    if cor < 0.75:
        a = 4
        img_br = cv2.convertScaleAbs(img, alpha=a, beta=0)
        all_images.append(img_br)
        cor_b, mac_b, coords_b, msg_b = get_macbeth_chart(img_br, ref_data)
        if cor_b > cor:
            cor, mac, coords, msg = cor_b, mac_b, coords_b, msg_b

    """
    In case macbeth chart is too small, take a selection of the image and
    attempt to locate macbeth chart within that. The scale increment is
    root 2
    """
    """
    These variables will be used to transform the found coordinates at smaller
    scales back into the original. If ii is still -1 after this section that
    means it was not successful
    """
    ii = -1
    w_best = 0
    h_best = 0
    d_best = 100
    """
    d_best records the scale of the best match. Macbeth charts are only looked
    for at one scale increment smaller than the current best match in order to avoid
    unecessarily searching for macbeth charts at small scales.
    If a macbeth chart ha already been found then set d_best to 0
    """
    if cor != 0:
        d_best = 0

    """
    scale 3/2 (approx root2)
    """
    if cor < 0.75:
        imgs = []
        """
        get size of image
        """
        shape = list(img.shape[:2])
        w, h = shape
        """
        set dimensions of the subselection and the step along each axis between
        selections
        """
        w_sel = int(2*w/3)
        h_sel = int(2*h/3)
        w_inc = int(w/6)
        h_inc = int(h/6)
        """
        for each subselection, look for a macbeth chart
        loop over this and any brightened up images that we made to increase the
        likelihood of success
        """
        for img_br in all_images:
            for i in range(3):
                for j in range(3):
                    w_s, h_s = i*w_inc, j*h_inc
                    img_sel = img_br[w_s:w_s+w_sel, h_s:h_s+h_sel]
                    cor_ij, mac_ij, coords_ij, msg_ij = get_macbeth_chart(img_sel, ref_data)
                    """
                    if the correlation is better than the best then record the
                    scale and current subselection at which macbeth chart was
                    found. Also record the coordinates, macbeth chart and message.
                    """
                    if cor_ij > cor:
                        cor = cor_ij
                        mac, coords, msg = mac_ij, coords_ij, msg_ij
                        ii, jj = i, j
                        w_best, h_best = w_inc, h_inc
                        d_best = 1

    """
    scale 2
    """
    if cor < 0.75:
        imgs = []
        shape = list(img.shape[:2])
        w, h = shape
        w_sel = int(w/2)
        h_sel = int(h/2)
        w_inc = int(w/8)
        h_inc = int(h/8)
        # Again, loop over any brightened up images as well
        for img_br in all_images:
            for i in range(5):
                for j in range(5):
                    w_s, h_s = i*w_inc, j*h_inc
                    img_sel = img_br[w_s:w_s+w_sel, h_s:h_s+h_sel]
                    cor_ij, mac_ij, coords_ij, msg_ij = get_macbeth_chart(img_sel, ref_data)
                    if cor_ij > cor:
                        cor = cor_ij
                        mac, coords, msg = mac_ij, coords_ij, msg_ij
                        ii, jj = i, j
                        w_best, h_best = w_inc, h_inc
                        d_best = 2

    """
    The following code checks for macbeth charts at even smaller scales. This
    slows the code down significantly and has therefore been omitted by default,
    however it is not unusably slow so might be useful if the macbeth chart
    is too small to be picked up to by the current subselections.
    Use this for macbeth charts with side lengths around 1/5 image dimensions
    (and smaller...?) it is, however, recommended that macbeth charts take up as
    large as possible a proportion of the image.
    """

    if small_chart:

        if cor < 0.75 and d_best > 1:
            imgs = []
            shape = list(img.shape[:2])
            w, h = shape
            w_sel = int(w/3)
            h_sel = int(h/3)
            w_inc = int(w/12)
            h_inc = int(h/12)
            for i in range(9):
                for j in range(9):
                    w_s, h_s = i*w_inc, j*h_inc
                    img_sel = img[w_s:w_s+w_sel, h_s:h_s+h_sel]
                    cor_ij, mac_ij, coords_ij, msg_ij = get_macbeth_chart(img_sel, ref_data)
                    if cor_ij > cor:
                        cor = cor_ij
                        mac, coords, msg = mac_ij, coords_ij, msg_ij
                        ii, jj = i, j
                        w_best, h_best = w_inc, h_inc
                        d_best = 3

        if cor < 0.75 and d_best > 2:
            imgs = []
            shape = list(img.shape[:2])
            w, h = shape
            w_sel = int(w/4)
            h_sel = int(h/4)
            w_inc = int(w/16)
            h_inc = int(h/16)
            for i in range(13):
                for j in range(13):
                    w_s, h_s = i*w_inc, j*h_inc
                    img_sel = img[w_s:w_s+w_sel, h_s:h_s+h_sel]
                    cor_ij, mac_ij, coords_ij, msg_ij = get_macbeth_chart(img_sel, ref_data)
                    if cor_ij > cor:
                        cor = cor_ij
                        mac, coords, msg = mac_ij, coords_ij, msg_ij
                        ii, jj = i, j
                        w_best, h_best = w_inc, h_inc

    """
    Transform coordinates from subselection to original image
    """
    if ii != -1:
        for a in range(len(coords)):
            for b in range(len(coords[a][0])):
                coords[a][0][b][1] += ii*w_best
                coords[a][0][b][0] += jj*h_best

    """
    initialise coords_fit variable
    """
    coords_fit = None
    # print('correlation: {}'.format(cor))
    """
    print error or success message
    """
    print(msg)
    Cam.log += '\n' + str(msg)
    if msg == success_msg:
        coords_fit = coords
        Cam.log += '\nMacbeth chart vertices:\n'
        Cam.log += '{}'.format(2*np.round(coords_fit[0][0]), 0)
        """
        if correlation is lower than 0.75 there may be a risk of macbeth chart
        corners not having been located properly. It might be worth running
        with show set to true to check where the macbeth chart centres have
        been located.
        """
        print('Confidence: {:.3f}'.format(cor))
        Cam.log += '\nConfidence: {:.3f}'.format(cor)
        if cor < 0.75:
            print('Caution: Low confidence guess!')
            Cam.log += 'WARNING: Low confidence guess!'
        # cv2.imshow('MacBeth', mac)
        # represent(mac, 'MacBeth chart')

    """
    extract data from coords_fit and plot on original image
    """
    if show and coords_fit is not None:
        copy = img.copy()
        verts = coords_fit[0][0]
        cents = coords_fit[1][0]

        """
        draw circles at vertices of macbeth chart
        """
        for vert in verts:
            p = tuple(np.round(vert).astype(np.int32))
            cv2.circle(copy, p, 10, 1, -1)
        """
        draw circles at centres of squares
        """
        for i in range(len(cents)):
            cent = cents[i]
            p = tuple(np.round(cent).astype(np.int32))
            """
            draw black circle on white square, white circle on black square an
            grey circle everywhere else.
            """
            if i == 3:
                cv2.circle(copy, p, 8, 0, -1)
            elif i == 23:
                cv2.circle(copy, p, 8, 1, -1)
            else:
                cv2.circle(copy, p, 8, 0.5, -1)
        copy, _ = reshape(copy, 400)
        represent(copy)

    return(coords_fit)


def get_macbeth_chart(img, ref_data):
    """
    function returns coordinates of macbeth chart vertices and square centres,
    along with an error/success message for debugging purposes. Additionally,
    it scores the match with a confidence value.

    Brief explanation of the macbeth chart locating algorithm:
    - Find rectangles within image
    - Take rectangles within percentage offset of median perimeter. The
        assumption is that these will be the macbeth squares
    - For each potential square, find the 24 possible macbeth centre locations
        that would produce a square in that location
    - Find clusters of potential macbeth chart centres to find the potential
        macbeth centres with the most votes, i.e. the most likely ones
    - For each potential macbeth centre, use the centres of the squares that
        voted for it to find macbeth chart corners
    - For each set of corners, transform the possible match into normalised
        space and correlate with a reference chart to evaluate the match
    - Select the highest correlation as the macbeth chart match, returning the
        correlation as the confidence score
    """

    """
    get reference macbeth chart data
    """
    (ref, ref_w, ref_h, ref_corns) = ref_data

    """
    the code will raise and catch a MacbethError in case of a problem, trying
    to give some likely reasons why the problem occred, hence the try/except
    """
    try:
        """
        obtain image, convert to grayscale and normalise
        """
        src = img
        src, factor = reshape(src, 200)
        original = src.copy()
        a = 125/np.average(src)
        src_norm = cv2.convertScaleAbs(src, alpha=a, beta=0)
        """
        This code checks if there are seperate colour channels. In the past the
        macbeth locator ran on jpgs and this makes it robust to different
        filetypes. Note that running it on a jpg has 4x the pixels of the
        average bayer channel so coordinates must be doubled.

        This is best done in img_load.py in the get_patches method. The
        coordinates and image width, height must be divided by two if the
        macbeth locator has been run on a demosaicked image.
        """
        if len(src_norm.shape) == 3:
            src_bw = cv2.cvtColor(src_norm, cv2.COLOR_BGR2GRAY)
        else:
            src_bw = src_norm
        original_bw = src_bw.copy()
        """
        obtain image edges
        """
        sigma = 2
        src_bw = cv2.GaussianBlur(src_bw, (0, 0), sigma)
        t1, t2 = 50, 100
        edges = cv2.Canny(src_bw, t1, t2)
        """
        dilate edges to prevent self-intersections in contours
        """
        k_size = 2
        kernel = np.ones((k_size, k_size))
        its = 1
        edges = cv2.dilate(edges, kernel, iterations=its)
        """
        find Contours in image
        """
        conts, _ = cv2.findContours(edges, cv2.RETR_TREE,
                                    cv2.CHAIN_APPROX_NONE)
        if len(conts) == 0:
            raise MacbethError(
                '\nWARNING: No macbeth chart found!'
                '\nNo contours found in image\n'
                'Possible problems:\n'
                '- Macbeth chart is too dark or bright\n'
                '- Macbeth chart is occluded\n'
            )
        """
        find quadrilateral contours
        """
        epsilon = 0.07
        conts_per = []
        for i in range(len(conts)):
            per = cv2.arcLength(conts[i], True)
            poly = cv2.approxPolyDP(conts[i], epsilon*per, True)
            if len(poly) == 4 and cv2.isContourConvex(poly):
                conts_per.append((poly, per))

        if len(conts_per) == 0:
            raise MacbethError(
                '\nWARNING: No macbeth chart found!'
                '\nNo quadrilateral contours found'
                '\nPossible problems:\n'
                '- Macbeth chart is too dark or bright\n'
                '- Macbeth chart is occluded\n'
                '- Macbeth chart is out of camera plane\n'
            )

        """
        sort contours by perimeter and get perimeters within percent of median
        """
        conts_per = sorted(conts_per, key=lambda x: x[1])
        med_per = conts_per[int(len(conts_per)/2)][1]
        side = med_per/4
        perc = 0.1
        med_low, med_high = med_per*(1-perc), med_per*(1+perc)
        squares = []
        for i in conts_per:
            if med_low <= i[1] and med_high >= i[1]:
                squares.append(i[0])

        """
        obtain coordinates of nomralised macbeth and squares
        """
        square_verts, mac_norm = get_square_verts(0.06)
        """
        for each square guess, find 24 possible macbeth chart centres
        """
        mac_mids = []
        squares_raw = []
        for i in range(len(squares)):
            square = squares[i]
            squares_raw.append(square)
            """
            convert quads to rotated rectangles. This is required as the
            'squares' are usually quite irregular quadrilaterls, so performing
            a transform would result in exaggerated warping and inaccurate
            macbeth chart centre placement
            """
            rect = cv2.minAreaRect(square)
            square = cv2.boxPoints(rect).astype(np.float32)
            """
            reorder vertices to prevent 'hourglass shape'
            """
            square = sorted(square, key=lambda x: x[0])
            square_1 = sorted(square[:2], key=lambda x: x[1])
            square_2 = sorted(square[2:], key=lambda x: -x[1])
            square = np.array(np.concatenate((square_1, square_2)), np.float32)
            square = np.reshape(square, (4, 2)).astype(np.float32)
            squares[i] = square
            """
            find 24 possible macbeth chart centres by trasnforming normalised
            macbeth square vertices onto candidate square vertices found in image
            """
            for j in range(len(square_verts)):
                verts = square_verts[j]
                p_mat = cv2.getPerspectiveTransform(verts, square)
                mac_guess = cv2.perspectiveTransform(mac_norm, p_mat)
                mac_guess = np.round(mac_guess).astype(np.int32)
                """
                keep only if candidate macbeth is within image border
                (deprecated)
                """
                in_border = True
                # for p in mac_guess[0]:
                #     pptest = cv2.pointPolygonTest(
                #         img_con,
                #         tuple(p),
                #         False
                #     )
                #     if pptest == -1:
                #         in_border = False
                #         break

                if in_border:
                    mac_mid = np.mean(mac_guess,
                                      axis=1)
                    mac_mids.append([mac_mid, (i, j)])

        if len(mac_mids) == 0:
            raise MacbethError(
                '\nWARNING: No macbeth chart found!'
                '\nNo possible macbeth charts found within image'
                '\nPossible problems:\n'
                '- Part of the macbeth chart is outside the image\n'
                '- Quadrilaterals in image background\n'
            )

        """
        reshape data
        """
        for i in range(len(mac_mids)):
            mac_mids[i][0] = mac_mids[i][0][0]

        """
        find where midpoints cluster to identify most likely macbeth centres
        """
        clustering = cluster.AgglomerativeClustering(
            n_clusters=None,
            compute_full_tree=True,
            distance_threshold=side*2
        )
        mac_mids_list = [x[0] for x in mac_mids]

        if len(mac_mids_list) == 1:
            """
            special case of only one valid centre found (probably not needed)
            """
            clus_list = []
            clus_list.append([mac_mids, len(mac_mids)])

        else:
            clustering.fit(mac_mids_list)
            # try:
            #     clustering.fit(mac_mids_list)
            # except RuntimeWarning as error:
            #     return(0, None, None, error)

            """
            create list of all clusters
            """
            clus_list = []
            if clustering.n_clusters_ > 1:
                for i in range(clustering.labels_.max()+1):
                    indices = [j for j, x in enumerate(clustering.labels_) if x == i]
                    clus = []
                    for index in indices:
                        clus.append(mac_mids[index])
                    clus_list.append([clus, len(clus)])
                clus_list.sort(key=lambda x: -x[1])

            elif clustering.n_clusters_ == 1:
                """
                special case of only one cluster found
                """
                # print('only 1 cluster')
                clus_list.append([mac_mids, len(mac_mids)])
            else:
                raise MacbethError(
                    '\nWARNING: No macebth chart found!'
                    '\nNo clusters found'
                    '\nPossible problems:\n'
                    '- NA\n'
                )

        """
        keep only clusters with enough votes
        """
        clus_len_max = clus_list[0][1]
        clus_tol = 0.7
        for i in range(len(clus_list)):
            if clus_list[i][1] < clus_len_max * clus_tol:
                clus_list = clus_list[:i]
                break
            cent = np.mean(clus_list[i][0], axis=0)[0]
            clus_list[i].append(cent)

        """
        represent most popular cluster centroids
        """
        # copy = original_bw.copy()
        # copy = cv2.cvtColor(copy, cv2.COLOR_GRAY2RGB)
        # copy = cv2.resize(copy, None, fx=2, fy=2)
        # for clus in clus_list:
        #     centroid = tuple(2*np.round(clus[2]).astype(np.int32))
        #     cv2.circle(copy, centroid, 7, (255, 0, 0), -1)
        #     cv2.circle(copy, centroid, 2, (0, 0, 255), -1)
        # represent(copy)

        """
        get centres of each normalised square
        """
        reference = get_square_centres(0.06)

        """
        for each possible macbeth chart, transform image into
        normalised space and find correlation with reference
        """
        max_cor = 0
        best_map = None
        best_fit = None
        best_cen_fit = None
        best_ref_mat = None

        for clus in clus_list:
            clus = clus[0]
            sq_cents = []
            ref_cents = []
            i_list = [p[1][0] for p in clus]
            for point in clus:
                i, j = point[1]
                """
                remove any square that voted for two different points within
                the same cluster. This causes the same point in the image to be
                mapped to two different reference square centres, resulting in
                a very distorted perspective transform since cv2.findHomography
                simply minimises error.
                This phenomenon is not particularly likely to occur due to the
                enforced distance threshold in the clustering fit but it is
                best to keep this in just in case.
                """
                if i_list.count(i) == 1:
                    square = squares_raw[i]
                    sq_cent = np.mean(square, axis=0)
                    ref_cent = reference[j]
                    sq_cents.append(sq_cent)
                    ref_cents.append(ref_cent)

                    """
                    At least four squares need to have voted for a centre in
                    order for a transform to be found
                    """
            if len(sq_cents) < 4:
                raise MacbethError(
                    '\nWARNING: No macbeth chart found!'
                    '\nNot enough squares found'
                    '\nPossible problems:\n'
                    '- Macbeth chart is occluded\n'
                    '- Macbeth chart is too dark or bright\n'
                )

            ref_cents = np.array(ref_cents)
            sq_cents = np.array(sq_cents)
            """
            find best fit transform from normalised centres to image
            """
            h_mat, mask = cv2.findHomography(ref_cents, sq_cents)
            if 'None' in str(type(h_mat)):
                raise MacbethError(
                    '\nERROR\n'
                )

            """
            transform normalised corners and centres into image space
            """
            mac_fit = cv2.perspectiveTransform(mac_norm, h_mat)
            mac_cen_fit = cv2.perspectiveTransform(np.array([reference]), h_mat)
            """
            transform located corners into reference space
            """
            ref_mat = cv2.getPerspectiveTransform(
                mac_fit,
                np.array([ref_corns])
            )
            map_to_ref = cv2.warpPerspective(
                original_bw, ref_mat,
                (ref_w, ref_h)
            )
            """
            normalise brigthness
            """
            a = 125/np.average(map_to_ref)
            map_to_ref = cv2.convertScaleAbs(map_to_ref, alpha=a, beta=0)
            """
            find correlation with bw reference macbeth
            """
            cor = correlate(map_to_ref, ref)
            """
            keep only if best correlation
            """
            if cor > max_cor:
                max_cor = cor
                best_map = map_to_ref
                best_fit = mac_fit
                best_cen_fit = mac_cen_fit
                best_ref_mat = ref_mat

            """
            rotate macbeth by pi and recorrelate in case macbeth chart is
            upside-down
            """
            mac_fit_inv = np.array(
                ([[mac_fit[0][2], mac_fit[0][3],
                  mac_fit[0][0], mac_fit[0][1]]])
            )
            mac_cen_fit_inv = np.flip(mac_cen_fit, axis=1)
            ref_mat = cv2.getPerspectiveTransform(
                mac_fit_inv,
                np.array([ref_corns])
            )
            map_to_ref = cv2.warpPerspective(
                original_bw, ref_mat,
                (ref_w, ref_h)
            )
            a = 125/np.average(map_to_ref)
            map_to_ref = cv2.convertScaleAbs(map_to_ref, alpha=a, beta=0)
            cor = correlate(map_to_ref, ref)
            if cor > max_cor:
                max_cor = cor
                best_map = map_to_ref
                best_fit = mac_fit_inv
                best_cen_fit = mac_cen_fit_inv
                best_ref_mat = ref_mat

        """
        Check best match is above threshold
        """
        cor_thresh = 0.6
        if max_cor < cor_thresh:
            raise MacbethError(
                '\nWARNING: Correlation too low'
                '\nPossible problems:\n'
                '- Bad lighting conditions\n'
                '- Macbeth chart is occluded\n'
                '- Background is too noisy\n'
                '- Macbeth chart is out of camera plane\n'
            )
            """
            Following code is mostly representation for debugging purposes
            """

        """
        draw macbeth corners and centres on image
        """
        copy = original.copy()
        copy = cv2.resize(original, None, fx=2, fy=2)
        # print('correlation = {}'.format(round(max_cor, 2)))
        for point in best_fit[0]:
            point = np.array(point, np.float32)
            point = tuple(2*np.round(point).astype(np.int32))
            cv2.circle(copy, point, 4, (255, 0, 0), -1)
        for point in best_cen_fit[0]:
            point = np.array(point, np.float32)
            point = tuple(2*np.round(point).astype(np.int32))
            cv2.circle(copy, point, 4, (0, 0, 255), -1)
            copy = copy.copy()
            cv2.circle(copy, point, 4, (0, 0, 255), -1)

        """
        represent coloured macbeth in reference space
        """
        best_map_col = cv2.warpPerspective(
            original, best_ref_mat, (ref_w, ref_h)
        )
        best_map_col = cv2.resize(
            best_map_col, None, fx=4, fy=4
        )
        a = 125/np.average(best_map_col)
        best_map_col_norm = cv2.convertScaleAbs(
            best_map_col, alpha=a, beta=0
        )
        # cv2.imshow('Macbeth', best_map_col)
        # represent(copy)

        """
        rescale coordinates to original image size
        """
        fit_coords = (best_fit/factor, best_cen_fit/factor)

        return(max_cor, best_map_col_norm, fit_coords, success_msg)

        """
    catch macbeth errors and continue with code
    """
    except MacbethError as error:
        return(0, None, None, error)