Thresholding

Description

Thresholding segments an image by comparing each pixel’s channel value against a threshold, producing pixels classified as either "foreground" or "background", or an image with saturated (clipped) intensities.

All functions described here operate on a per-channel basis, so they apply equally to grayscale and multi-channel (e.g. RGB) images, and accept a threshold_direction:

  • threshold_direction::regular treats values greater than the threshold as foreground.

  • threshold_direction::inverse treats values less than or equal to the threshold as foreground.

Binary threshold

threshold_binary sets each pixel to max_value or 0, depending on whether it lies on the foreground or background side of threshold_value (swapped when direction is inverse). If max_value is omitted, it defaults to the maximum value representable by the destination channel type.

#include <boost/gil/image_processing/threshold.hpp>

// values above 150 become 255, others become 0
threshold_binary(view(img), view(img_out), 150, 255);

Truncating threshold

threshold_truncate leaves foreground pixels either clipped to threshold_value or set to 0, depending on threshold_truncate_mode:

  • threshold_truncate_mode::threshold clips foreground pixels to threshold_value and leaves background pixels unchanged.

  • threshold_truncate_mode::zero zeroes background pixels and leaves foreground pixels unchanged.

// values above 150 are clipped down to 150, others are left unchanged
threshold_truncate(view(img), view(img_out), 150, threshold_truncate_mode::threshold);

Optimal threshold

threshold_optimal computes the threshold value itself instead of taking one as a parameter. The only method currently implemented, threshold_optimal_value::otsu, picks the threshold that minimizes intra-class intensity variance between the two resulting classes, then applies it as a binary threshold.

threshold_optimal(view(img), view(img_out), threshold_optimal_value::otsu);

Adaptive threshold

Unlike the fixed thresholds above, threshold_adaptive computes a different threshold for each pixel, based on the pixels in its local neighborhood. This copes better than a single, fixed threshold with images where illumination varies across the frame.

For each pixel, the local threshold is the value of a blurred version of the image at that pixel (optionally offset by constant), where the blur is computed by convolving with a kernel_size x kernel_size kernel chosen by threshold_adaptive_method:

  • threshold_adaptive_method::mean averages the neighborhood uniformly.

  • threshold_adaptive_method::gaussian weights the neighborhood with a Gaussian kernel, giving nearby pixels more influence than distant ones.

kernel_size must be odd. If max_value is omitted, it defaults to the maximum value representable by the destination channel type.

#include <boost/gil/image_processing/threshold.hpp>

// mean-adaptive threshold with an 11x11 neighborhood, offset by 2
threshold_adaptive(view(img), view(img_out), 11, threshold_adaptive_method::mean,
    threshold_direction::regular, 2);

// gaussian-adaptive threshold with a 7x7 neighborhood, offset by 2
threshold_adaptive(view(img), view(img_out), 7, threshold_adaptive_method::gaussian,
    threshold_direction::regular, 2);

Demo

See the full examples at example/threshold.cpp and example/adaptive_threshold.cpp.

Edit this Page