Draw on the canvas to understand how images are represented as matrices and apply basic filters.
At its core, a digital grayscale pic is nothing more than a grid of counts—a matrix. Each count, called a pixel, represents the brightness at that set point, often ranging from 0 (black) to 255 (white). This interactive bridges the gap between the visual pic you draw and the underlying numerical data that a computer really "sees" and manipulates. Each pic step — blurring, sharpening, edge detection, thresholding — is a mathematical step on that matrix. A blur replaces each pixel with the average of its neighbours. An edge detector computes the gap between next pixels. Convolutions generalise both: slide a small kernel over the pic, multiply item-wise. And sum the results into a new pixel. A colour pic adds two more channels (red, green, blue), turning the matrix into a 3D tensor. But the steps are the same per channel. Turning colour to grayscale collapses three channels into one, often with the luminance formula 0.299R + 0.587G + 0.114B. It weights green most heavily since human eyes are most sensitive to it. This page lets you draw on a canvas, shows the underlying pixel matrix in real time. And shows how basic steps (invert, threshold, blur) transform the counts.
Draw on the canvas to understand how images are represented as matrices and apply basic filters.
From visual drawings to numerical matrices—the foundation of computer vision.
At its core, a digital grayscale pic is nothing more than a grid of counts—a matrix. Each count, called a pixel, represents the brightness at that set point, often ranging from 0 (black) to 255 (white). This interactive bridges the gap between the visual pic you draw and the underlying numerical data that a computer really "sees" and manipulates. Each pic step — blurring, sharpening, edge detection, thresholding — is a mathematical step on that matrix. A blur replaces each pixel with the average of its neighbours. An edge detector computes the gap between next pixels. Convolutions generalise both: slide a small kernel over the pic, multiply item-wise. And sum the results into a new pixel. A colour pic adds two more channels (red, green, blue), turning the matrix into a 3D tensor. But the steps are the same per channel. Turning colour to grayscale collapses three channels into one, often with the luminance formula 0.299R + 0.587G + 0.114B. It weights green most heavily since human eyes are most sensitive to it. This page lets you draw on a canvas, shows the underlying pixel matrix in real time. And shows how basic steps (invert, threshold, blur) transform the counts.
The most basic concept in all of computer vision is that pics are just data. Each step, from a simple filter to a complex neural network, is a mathematical function done on this matrix of pixel vals. By grasp this relationship, you gain ability to understand how pic handling algorithms work.
As you draw on the canvas, you are right changing the counts in the matrix on the right. A bright stroke sets pixel vals to 255, while the empty canvas is all zeros. Each filter you apply is just a rule that transforms these counts.
A greyscale pic stores a single brightness value per pixel, conventionally 0 for black and 255 for white in 8-bit pics. A colour pic stores three.
That reduction has a direct cost and a direct gain. A 1920×1080 pic is 2 million vals in greyscale and 6 million in colour. So turning cuts memory and computation by two thirds. What you lose is each distinction that depends on hue: a red apple and a green apple of the same brightness turn identical.
The change is not a plain average. Human vision is far more sensitive to green than to blue. So the standard weighting reflects perceived brightness:
Y = 0.299R + 0.587G + 0.114B
A plain (R+G+B)/3 makes blues look too light and greens too dark. Both are one line of code. The weighted version is the one that matches what people see.
Thresholding turns greyscale into binary — each pixel turns black or white:
import cv2
_, binary = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
_, otsu = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
adaptive = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, blockSize=11, C=2)A fixed threshold fails as soon as lighting varies. Otsu's way chooses the threshold automatically by finding the value that best separates the histogram into two groups. Adaptive thresholding computes a other threshold for each neighbourhood. And is what you want for a photograph of a document with a shadow across it.
Histogram equalisation redistributes brightness vals to use the full range, revealing detail in pics that are too dark or too flat. CLAHE — the adaptive version, applied in tiles — avoids the over-amplified noise that global equalisation gives in now-good regions.
Morphological steps work on binary pics: erosion shrinks white regions, dilation grows them, opening (erode then dilate) removes small specks. And closing (dilate then erode) fills small holes. They are the standard cleanup after thresholding.
| Task | Greyscale enough? |
|---|---|
| OCR and document processing | Yes — text is defined by shape |
| Edge and corner detection | Yes — brightness gradients carry it |
| Medical X-ray and CT | Yes — the data is single-channel already |
| Face detection (classical) | Yes — Haar cascades use intensity patterns |
| Fruit ripeness, traffic lights | No — colour is the signal |
| Skin lesion analysis | No — hue carries diagnostic information |
| General object classification | Usually not — colour helps materially |
The rule of thumb: if a human could do the task from a black-and-white photograph, greyscale is probably enough. If they would ask to see it in colour, it is not.
Draw a simple diagonal line on the canvas. Watch the corresponding cells in the matrix light up with vals like 255, 120, and 80. Notice how the "brush" has a soft edge, creating intermediate gray vals, not just pure white.
Click the Invert Colors button. Observe the matrix. Each pixel value p has been replaced by 255 - p. The black background (0) turns white (255). And your white drawing (255) turns black (0). This is a simple but powerful pixel-wise step.
Draw a shape and then click Apply Threshold with the slider at 128. All pixels with a value greater than 128 turn pure white (255). And all others turn pure black (0). This is called binarization. And it's a common step to simplify an pic and isolate objects of interest.
Clear the canvas and draw a single, sharp dot. Now click Gaussian Blur. Look at the matrix. The single high-value pixel has been "spread out" to its neighbors, creating a smooth gradient. This is reached by averaging each pixel with its surrounding pixels. It is key for reducing noise in pics.
Draw a solid square. Click Edge Detection. Notice that just the borders of your square remain as bright pixels in the matrix. The solid interior turns black (0). This filter calculates the gap between neighboring pixels. Where the gap is large (at an edge), the output value is high.
Mastering this basic concept—that pics are just grids of counts—is the key to grasp all of computer vision.
Feeding greyscale pics to a CNN changes just one thing: the input has one channel instead of three. So the first layer's filters have a third as many weights. All after that is unchanged.
Two real points arise when mixing greyscale data with pretrained nets.
Pretrained networks expect three channels. A ResNet trained on ImageNet has a first layer shaped for RGB. The usual fix is to replicate the greyscale channel three times. It works and wastes a little computation. The better fix is to sum the pretrained first-layer weights across the channel dimension, giving a single-channel filter that behaves identically on grey input.
Colour augmentation stops applying. Hue and saturation jitter are meaningless on single-channel data. Brightness and contrast augmentation still apply, and matter more.
Whether to discard colour is worth testing rather than assuming. On many datasets the accuracy gap is small and the speed gain is real. On others — anything where category correlates with colour — it is a large loss.
Most consumer pics use 8 bits per channel: 256 levels. That is enough for a photograph on a screen and not enough for a few serious applications.
Medical imaging uses 12 or 16 bits, giving 4,096 or 65,536 levels, since subtle tissue gaps occupy a narrow band that 8 bits would flatten. Scientific and astronomical imaging is close. Satellite data often carries more.
When working with such data, two habits matter: do not casually turn to 8 bits for convenience. And normalise with the actual data range rather than assuming 0–255. A 16-bit medical pic divided by 255 gives vals in the thousands. And a network fed those will not train.
Turning to greyscale looks like the most trivial step in vision. And the obvious way to do it is wrong. This shows what the standard weights are for, what they cost you. And the one place the naive average is really the right answer.
Building on the same setup, the next block shows the left cases.
The next block covers the left steps.
The next block continues with the left demonstrations.
Why not just average the channels? Since human brightness perception is not uniform across colours. The weighted formula matches perception; the average does not.
Does greyscale change lose information? Yes, irreversibly — many colours map to the same grey value.
Is greyscale faster to process? Roughly three times less data. And the first convolution layer is a third the size. Later layers are unaffected.
Should I turn before or after augmentation? Before, if you are not with colour augmentation. The pipeline is then cheaper throughout.
What is the gap between greyscale and binary? Greyscale has a range of intensities. Binary has just two vals, produced by thresholding.
When does adaptive thresholding beat Otsu? Whenever illumination varies across the pic — a photographed page with a shadow is the canonical case.
Say the answer out loud before you reveal it — recalling it is what makes it stick, and rereading it is not.
What is meant by “Pixel-wise steps” here?
What is meant by “Neighborhood steps” here?
See a grayscale image for what it is: a grid of brightness numbers you can filter, threshold and edit directly.