CV 418 — Concepts Made Simple (understand everything)

What is a digital image, in one sentence?

A grid of numbers for brightness — the camera turns real-world light into numbers a computer can store, like filling in a big spreadsheet of dots.

In the eye, which part does MOST of the focusing of light?

The cornea — the clear front window bends most of the incoming light, doing the heavy lifting before anything else.

What does the iris do in the eye?

It's the aperture — it opens and closes the pupil to control how much light gets in, like blinds on a window.

What is the retina?

The screen at the back of the eye that holds the light-sensors and turns light into nerve signals — like the film or sensor in a camera.

Retina vs optic nerve: which one converts light, and which one carries the signal?

The retina converts light into signals; the optic nerve just carries them to the brain — one is the sensor, the other is the cable.

Which eye receptors handle color and sharp detail in bright light?

Cones — they give you color and crisp detail, and they're packed in the center; there are only about 6-7 million per eye.

Which eye receptors handle dim, night vision and see only gray?

Rods — they work in the dark, sit off to the sides, see no color, and there are way more of them (about 75-150 million per eye).

What is the eye's 'blind spot'?

The exact spot where the optic nerve leaves the eye — it has no receptors there, so nothing lands on that patch.

A camera focuses by moving its lens. How does the EYE focus?

By changing the lens SHAPE — it squishes the lens fatter or flatter instead of moving it, since it can't slide the lens back and forth.

Is visible light a big or tiny part of all the light (the EM spectrum) out there?

A tiny sliver — the full spectrum runs from radio waves to gamma rays, and our eyes only catch a thin band in the middle.

Wavelength and frequency: if one goes up, what does the other do?

The other goes down — they trade off, so a higher frequency always means a shorter wavelength (they're linked by the fixed speed of light).

Your eye handles a huge range of light — but all at once, or by adjusting?

By adjusting — it 'brightness adapts,' shifting its sensitivity to whatever light level you're in, like your eyes adjusting when you enter a dark room.

Does doubling the light make a scene feel twice as bright?

No — your eyes react to RATIOS, not raw amounts. Going from 1 lamp to 2 feels like a big jump, but 100 lamps to 200 feels tiny, even though both doubled. So doubling reads as just one small step up, not double.

If someone can notice a really tiny change in brightness, is that good or bad eyesight?

Good — spotting tiny changes means sharp eyes. (The 'Weber ratio' is just this idea turned into a number: a smaller number means you catch finer changes, so small ratio = sharp eyes.)

What are Mach bands?

Fake bright/dark stripes your brain paints right at the edge between two shades — they aren't really in the picture; your brain over-reacts at borders.

Why does the same gray square look darker on white and lighter on black (simultaneous contrast)?

Because your brain judges brightness RELATIVE to the surroundings, not on its own — the background changes how the same gray reads.

'Sampling' means chopping the picture into a grid of dots. What does it decide?

How many pixels you get — more samples means more pixels; too few and the image looks blocky (pixelated).

'Quantization' means rounding each pixel's brightness to a fixed set of values. What does it decide?

How many brightness levels you get — like a dimmer switch with only a few settings; too few levels gives ugly banding in smooth areas.

With 8 bits per pixel, how many brightness levels do you get?

256 (values 0 to 255) — it's 2 multiplied by itself 8 times, NOT 2 times 8; each extra bit doubles the choices.

In image coordinates, where is the origin (0,0)?

The TOP-LEFT corner — and y goes DOWN, unlike the math graphs where the origin is bottom-left and y goes up.

How do you get the storage size of an image in bits?

Multiply rows x columns x bits-per-pixel — count every pixel, then count the bits each one needs; divide by 8 for bytes.

Surprisingly, which image is hurt MOST by cutting resolution — a smooth face or a busy crowd?

The smooth face — low-detail images suffer most, because your eye easily notices blockiness in smooth areas; busy crowds hide it.

What do we call the straight, as-the-crow-flies distance between two pixels?

Euclidean distance — the distance you'd measure with a ruler held diagonally. If two dots sit 3 apart sideways and 4 apart up, the slanted line between them is exactly 5.

City-block (D4) distance only allows up/down/left/right steps. For a move of 3 across and 4 up, what is it?

7 — you just add 3 and 4, because you can't cut the diagonal; it's like walking city blocks in a grid, no shortcuts.

Chessboard (D8) distance lets you move like a chess king. For a move of 3 across and 4 up, what is it?

4 — you take the BIGGER of the two (3 or 4), since a king slides diagonally to cover both directions at once.

What does '4-adjacency' mean — which pixels count as a pixel's neighbors?

Its 4 touching neighbors — up, down, left, right, but NO diagonals; 8-adjacency adds the 4 diagonal corners for 8 in total.

What is a 'pixel'?

The tiniest dot that makes up a picture — a screen or photo is really a huge grid of these little squares, each holding one brightness or color value. Zoom in far enough and you can see them.

When we talk about a pixel's 'value' or 'intensity', what do we mean?

Just how bright that dot is, written as a number. Bigger number means brighter. In a normal image that number runs from 0 (black) up to 255 (white).

What is a digital image, really?

A grid of numbers — like a giant spreadsheet where each little square holds a number for how bright that dot is.

Each single number in an image grid stands for one what?

One pixel — one tiny dot of the picture, holding its brightness value.

In a black-and-white image, what does the number 100 mean?

A medium gray — since brightness runs 0 (black) up to 255 (white), 100 sits in the middle.

When the computer loads an image, what kind of thing is it?

A NumPy array — basically a table of numbers Python can do math on, not a photo file or a list.

Since an image is just numbers, what happens if you add 50 to every pixel?

The whole picture gets brighter — bumping every brightness number up 50 is like nudging a dimmer switch up.

A plain gray image is a grid with how many numbers per pixel?

One — a single brightness number per dot, arranged in rows and columns (height by width).

A color image stores how many numbers per pixel?

Three — one each for the blue, green, and red amounts, like a stack of three see-through layers.

In an image's shape, which comes first — height or width?

Height (the rows) comes first, then width, then color layers — the opposite of the (x, y) order from math class.

A 200-tall, 300-wide color photo has what shape?

(200, 300, 3) — 200 rows, 300 columns, and 3 color layers; never flip the first two.

In that shape, what is the trailing 3?

The number of color layers (channels), not a color value — it just says 'this pixel has 3 numbers.'

OpenCV stores color in what surprising order?

BGR — Blue, Green, Red, backwards from the usual RGB. It's just how OpenCV was built.

In OpenCV, what color is the pixel [255, 0, 0]?

Pure blue — because blue is listed first in BGR, not red like you'd expect.

Colors look swapped when you show an OpenCV image with a plotting tool. Why?

The plotter expects RGB but got BGR — you must convert first, or reds and blues trade places.

What does 'uint8' mean for a pixel number?

A whole number from 0 to 255 with no minus sign — the standard size for one brightness value (that's 256 possible values).

What is the biggest value a normal pixel can hold?

255 — numbers can't go above that or below 0 in a standard image.

With plain '+', what does 250 + 10 become in a pixel?

4 — it wraps around past the max like an odometer rolling over from 255 back to 0.

What does cv2.add do differently with 250 + 10?

It stops at 255 — instead of wrapping, it clips at the top, like a cup that just won't overflow.

'Bit depth' controls how many of what?

How many different colors or shades are possible — more bits means more shades, doubling each time you add a bit.

How many colors does 24-bit 'true color' give?

About 16.7 million — that's 256 shades of blue times green times red, way more than the eye can tell apart.

Is bit depth the same as resolution?

No — bit depth is how many colors, resolution is how many pixels; two totally separate ideas.

How do you find how many bytes a gray image takes?

Multiply height by width — one byte per pixel, so a 200 by 300 gray image is 60,000 bytes.

How many bytes does a color image take?

Height times width times 3 — three color layers means three bytes per pixel. Never multiply by the number of colors.

What is a 'binary' image?

A picture where every pixel is only pure black (0) or pure white (255) — nothing in between, like a stencil.

In an indexed image (like a GIF), what does each pixel store?

A lookup number pointing to one color in a fixed palette of 256 — like painting by numbers instead of storing full colors, which saves space.

To read one pixel you write image[y, x]. Which comes first?

The row (y, height) comes first, then the column (x) — again the reverse of math's (x, y).

What does the alpha channel control?

Transparency — 0 is fully see-through and 255 is fully solid. It's about see-through-ness, not brightness or a 4th color.

What is a 'bit', in plain words?

The smallest scrap of computer memory — a single on/off switch, a 0 or a 1. Everything a computer stores, including images, is built out of piles of these tiny switches.

What does 'resolution' mean?

How many pixels a picture is made of, usually given as width-by-height (like 1920 by 1080). More pixels means a bigger, more detailed image; fewer pixels means it looks blocky.

In a color photo, how many numbers describe one dot (pixel)?

Three — one each for Red, Green, Blue. Every color you see is just a recipe of those three amounts.

Each of the Red, Green, Blue numbers can go from 0 up to what?

255 — 0 means none of that light, 255 means full blast of it.

What do you get when you mix full Red, Green, and Blue light together?

White — mixing colored LIGHT adds up to white, which is why RGB is called 'additive.'

OpenCV loads a color image with the three numbers in what order?

Blue, Green, Red (BGR) — it's backwards from the usual RGB, and that's the classic trap.

You show an OpenCV image directly and the reds look blue. What went wrong?

You forgot to flip BGR to RGB — the plotting tool expects RGB, so swap it first with cvtColor.

Printer ink mixes toward black instead of white. What is that kind of color mixing called?

Subtractive (CMY) — ink SOAKS UP light, so more ink means darker, the opposite of glowing screens.

If a pixel's Red is 200, what is its Cyan ink amount?

55 — you just do 255 minus 200. Cyan is 'everything except red,' so more red means less cyan.

In a grayscale image, how many numbers describe one pixel?

Just one — pure brightness, 0 is black and 255 is white, with no color at all.

When turning color into gray, is it a plain average of Red, Green, Blue?

No — it's a weighted mix where green counts most and blue least, because our eyes see green best.

A picture keeps only the red things and blacks out the rest. Is that grayscale?

No — grayscale removes ALL color. Keeping one color is color filtering (done in HSV), not grayscale.

A binary image only allows which two pixel values?

0 and 255 (black and white) — nothing in between. It's the extreme, simplest version of an image.

What do the three letters in HSV stand for?

Hue (which color), Saturation (how vivid), Value (how bright) — it separates 'what color' from 'how bright.'

Why do people pick colors in HSV instead of RGB when lighting changes?

Because the Hue (which color) barely shifts when it gets brighter or darker, so one hue reliably catches the color.

In OpenCV, the Hue number runs from 0 to what? (the famous trap)

0 to 179 — the real 0-360 degrees is halved so it fits in one byte. Saturation and Value still go to 255.

You want to grab only the blue objects in a photo. What does inRange give you?

A black-and-white mask — white where the pixel's color is inside your blue range, black everywhere else.

After making a mask, how do you keep only the blue pixels and black out the rest?

bitwise_and with the mask — it's like a stencil that only lets the chosen pixels show through.

Which color space is 'perceptually uniform,' best for measuring color differences?

LAB — equal number-changes look like equal color-changes to the eye, so it's great for matching colors.

Which color space cleanly splits brightness from color and is used to shrink JPEG/video files?

YCrCb — 'Y' is pure brightness, and separating it lets compression throw away color info we barely notice.

Name the four ways you can move or reshape an image on a grid.

Translate (slide), Rotate (spin), Scale (resize), Shear (slant) — the four basic geometric moves.

You scale a 500-pixel side by a factor of 1.5. What's the new length?

750 — you just multiply the old size by the factor. Scaling is plain multiplication.

Why is sliding a picture sideways handled a bit differently from spinning or resizing it?

Because sliding means ADDING the same shift to every dot, while spinning and resizing are done by MULTIPLYING. The computer uses a small bookkeeping trick so a single step can do both adding and multiplying at once — that way all the moves can work the same way.

After rotating, a pixel lands at spot (12.3, 45.6). Why is that a problem?

Real pixels only sit at whole-number spots, so there's nothing exactly at 12.3 — you must GUESS the value there.

Guessing the color for an in-between spot has a name. What is it?

Interpolation — literally estimating a value that sits between the ones you actually know.

The simplest guess just grabs the single closest pixel. What's that method called?

Nearest-Neighbor — fastest, but edges come out jagged and blocky because it copies one pixel wholesale.

Bilinear interpolation blends how many surrounding pixels?

4 (a 2x2 box) — a distance-weighted average of the four corners, giving a smoother, softer result.

Rank these by how many pixels they blend: Nearest, Bilinear, Bicubic, Lanczos.

1, then 4, then 16, then ~64 — more pixels means better quality but slower, so Lanczos is best but priciest.

What is this whole chapter really about?

Making pictures clearer by changing how bright or dark each pixel looks — like fiddling with the brightness and contrast knobs on a TV.

In an 8-bit image, what does 0 mean and what does 255 mean?

0 is pure black, 255 is pure white, and everything between is a shade of gray — like a dimmer switch from off to full.

A 'point transform' looks at how many pixels at once to decide a pixel's new brightness?

Just one — its own. It never peeks at the neighbors, unlike blurring, which mixes nearby pixels together.

A recolor rule is written s = T(r). What are r and s?

r is the old brightness, s is the new brightness — T is just the recipe that turns old into new.

What does a 'negative' transform do to an image? (rule: new = 255 minus old)

It flips dark and bright, so black becomes white — exactly like an old photo negative, great for revealing detail in dark areas.

What is a Look-Up Table (LUT) and why is it a speed trick?

You work out the answer for all 256 brightnesses once, then just look each pixel up — like a times-table you memorize instead of redoing the math.

What does the 'log' transform do to brightness?

It spreads out the dark values and squishes the bright ones — it rescues detail hiding in the shadows of a super-bright scene.

Gamma below 1 makes an image do what — brighter or darker?

Brighter — a small gamma (like a square root) lifts the dark tones up. Memory hook: 'a fraction power brightens.'

Gamma above 1 makes an image do what?

Darker — the opposite of a small gamma. People flip this constantly, so pin it: big gamma equals darker.

What does a 'contrast stretch' do?

It takes the darkest and brightest pixels the image actually has and stretches them to fill the full 0-to-255 range — like pulling a dull, flat photo taut so it pops.

What shape does a 'sigmoid' transform give brightness, and what does it boost?

An S-shape that boosts the middle grays while softening the very dark and very bright ends — it makes mid-tones pop without blowing out the extremes.

What is a 'piecewise-linear' transform?

A transform built from a few straight-line pieces joined together, each with its own steepness — like a rule that treats dark, medium, and bright ranges differently.

What is intensity (gray-level) slicing?

You pick one band of brightness and highlight just it — like a highlighter that lights up only the medium-gray stuff and ignores the rest.

Thresholding with BINARY: a pixel above the cutoff becomes what?

Pure white (255); everything else goes black — it splits the image into just two colors, like a stark black-and-white stamp.

When would you use BINARY_INV instead of plain BINARY?

When your object is darker than its background — it flips the rule so the dark thing turns white instead.

With a cutoff of 128, TRUNC turns a bright pixel of 200 into what?

128 — TRUNC caps everything at the cutoff instead of pushing it to 255, handy for taming glaring bright spots like headlights.

TOZERO keeps the bright side; what does TOZERO_INV keep?

The dark side — it keeps dark things like cracks and fingerprints and zeroes out the bright stuff, the mirror image of TOZERO.

What is a histogram of an image?

A tally of how many pixels have each brightness — like counting how many people are each height and drawing a bar chart of it.