← All demos

Computer vision

Blob detection

synthetic binary image, morphological cleanup, and connected-component (blob) analysis.

catalog/qu_image_blobs.qu

Builds a small grayscale image entirely in Qu (a few rectangular "blobs" painted onto a noisy background), thresholds it with Otsu's method, cleans up salt-and-pepper noise with an opening, then labels and measures the surviving blobs -- a minimal machine-vision pipeline with no image file on disk needed.

What it exercises

• image_from_matrix(mat) -- build a Value::Image straight from a matrix
• otsu_threshold / threshold -- automatic binarization
• imopen(img, radius) -- erosion+dilation to remove speckle noise
• label_blobs / bwlabel -- connected-component labeling, model kind "blobs"
• blob_stats / regionprops -- per-blob area/centroid/bounding box

How it works

W = 64
H = 48
canvas = 20 * ones(H, W)

canvas[5:14, 8:19] = 200 * ones(10, 12)
canvas[25:39, 5:17] = 180 * ones(15, 13)
canvas[10:29, 40:54] = 220 * ones(20, 15)

canvas = canvas + 15 * randn(H, W)
img = image_from_matrix(canvas)
print("Synthetic image: {img.width}x{img.height}")

t = otsu_threshold(img)
binary = threshold(img, t)
print("Otsu threshold = {t:.1f}")

clean = imopen(binary, 1)

labeled = label_blobs(clean)
print("Blobs found: {labeled.count}")

stats = blob_stats(labeled)
for i = 0 to length(stats) - 1
    b = stats[i]
    print("blob {b.label}: area={b.area} centroid=({b.centroid_x:.1f}, {b.centroid_y:.1f}) bbox={b.bbox_width}x{b.bbox_height}")
end for

subplot(1, 3, 1)
imshow(img)
title("Noisy source")

subplot(1, 3, 2)
imshow(binary)
title("Otsu + threshold")

subplot(1, 3, 3)
imshow(clean)
title("After imopen")

show plot

Results

Synthetic image: 64x48
Otsu threshold = 73.7
Blobs found: 3
blob 1: area=120 centroid=(13.5, 9.5) bbox=12x10
blob 2: area=300 centroid=(47.0, 19.5) bbox=15x20
blob 3: area=195 centroid=(11.0, 32.0) bbox=13x15

The whole program

Runnable as it stands: qu run catalog/qu_image_blobs.qu

# qu_image_blobs.qu -- synthetic binary image, morphological cleanup, and

# connected-component (blob) analysis.

#

# Builds a small grayscale image entirely in Qu (a few rectangular "blobs"

# painted onto a noisy background), thresholds it with Otsu's method, cleans

# up salt-and-pepper noise with an opening, then labels and measures the

# surviving blobs -- a minimal machine-vision pipeline with no image file on

# disk needed.

#

# Constructs exercised (all real, verified-working builtins):

#   image_from_matrix(mat)      -- build a Value::Image straight from a matrix

#   otsu_threshold / threshold  -- automatic binarization

#   imopen(img, radius)         -- erosion+dilation to remove speckle noise

#   label_blobs / bwlabel       -- connected-component labeling, model kind "blobs"

#   blob_stats / regionprops    -- per-blob area/centroid/bounding box


backend auto

#%% Paint a few rectangular blobs onto a noisy background

W = 64
H = 48
canvas = 20 * ones(H, W)

canvas[5:14, 8:19] = 200 * ones(10, 12)
canvas[25:39, 5:17] = 180 * ones(15, 13)
canvas[10:29, 40:54] = 220 * ones(20, 15)

canvas = canvas + 15 * randn(H, W)
img = image_from_matrix(canvas)
print("Synthetic image: {img.width}x{img.height}")

#%% Otsu binarization, then morphological cleanup

t = otsu_threshold(img)
binary = threshold(img, t)
print("Otsu threshold = {t:.1f}")

clean = imopen(binary, 1)

#%% Connected-component labeling and per-blob measurements

labeled = label_blobs(clean)
print("Blobs found: {labeled.count}")

stats = blob_stats(labeled)
for i = 0 to length(stats) - 1
    b = stats[i]
    print("blob {b.label}: area={b.area} centroid=({b.centroid_x:.1f}, {b.centroid_y:.1f}) bbox={b.bbox_width}x{b.bbox_height}")
end for

#%% Show the pipeline stages

subplot(1, 3, 1)
imshow(img)
title("Noisy source")

subplot(1, 3, 2)
imshow(binary)
title("Otsu + threshold")

subplot(1, 3, 3)
imshow(clean)
title("After imopen")

show plot