Which cv2.resize interpolation flag should you use when downscaling, and why?
answer
- direction decides the flag
- shrinking is averaging, not sampling
- most source pixels never get read
- moire on fine texture
- masks never get averaged
basics
~20 sUse cv2.INTER_AREA. It averages over the source pixel area that maps to each output pixel, so detail is integrated rather than skipped. The default cv2.INTER_LINEAR samples only a few points and produces aliasing and moire on large reductions.
solid answer
~40 sFor shrinking, pass `interpolation=cv2.INTER_AREA` to `cv2.resize`. It resamples using the pixel-area relation: each destination pixel is the average of the whole source region that maps onto it, which is the antialiasing you need when many source pixels collapse into one. The default, `cv2.INTER_LINEAR`, interpolates from only the four nearest source pixels regardless of the scale factor, so on a 10x reduction it is effectively sampling and most of the image never contributes — you get aliasing, moire on fine textures, and unstable results frame to frame. For enlarging, `INTER_AREA` behaves roughly like nearest-neighbour, so switch to `cv2.INTER_LINEAR` (fast) or `cv2.INTER_CUBIC` / `cv2.INTER_LANCZOS4` (slower, sharper). For a label or segmentation mask use `cv2.INTER_NEAREST` in either direction, because averaging class ids invents ids that mean nothing.
code
python · 11 linesimport cv2, numpy as np
# fine vertical stripes: the classic aliasing target
img = np.zeros((1000, 1000), np.uint8)
img[:, ::2] = 255
area = cv2.resize(img, (100, 100), interpolation=cv2.INTER_AREA)
linear = cv2.resize(img, (100, 100), interpolation=cv2.INTER_LINEAR)
print("area std:", round(float(area.std()), 2)) # near 0: stripes averaged to grey
print("linear std:", round(float(linear.std()), 2)) # high: aliased pattern survivesgo deeper
Memorise the two rules: INTER_AREA when making an image smaller, INTER_LINEAR or INTER_CUBIC when making it bigger, and remember dsize is (width, height).
Explain why — decimation needs averaging over the mapped source region while interpolation estimates between samples — and know the mask exception where only INTER_NEAREST is safe.
Watch for the cross-library seam: state that preprocessing at inference must match training, and be able to point at aliasing artefacts in a pipeline as the cause of unstable downstream results.
Treat resize policy as part of the data contract between capture, training and serving, and make it explicit and tested rather than a per-call flag anyone can change.
## What resizing actually has to do `cv2.resize(src, dsize, fx=..., fy=..., interpolation=...)` builds an output grid and has to decide what value each output pixel takes. There are two very different regimes: - **Upscaling**: output pixels are denser than input pixels. Each output pixel falls between known samples, and the job is *interpolation* — estimate a plausible value between neighbours. - **Downscaling**: many input pixels map into one output pixel. The job is *decimation*, and the correct operation is to integrate (average) over the region, not to interpolate at a point. Sampling a signal more sparsely than its detail warrants folds high frequencies down into fake low frequencies — aliasing — which is what makes a shrunken picket fence or striped shirt shimmer with moire patterns. That asymmetry is why one flag is not right for both directions. ## The flags - `cv2.INTER_NEAREST` — copies the closest source pixel. Fast, blocky, invents no new values. The right choice for label maps, index images and any data where the numbers are categories rather than intensities. `cv2.INTER_NEAREST_EXACT` is a variant with different rounding at exact half-pixel positions. - `cv2.INTER_LINEAR` — bilinear from the 4 nearest pixels. This is the **default**, a good general choice for upscaling and mild changes, and the wrong choice for big reductions. `cv2.INTER_LINEAR_EXACT` uses a bit-exact computation. - `cv2.INTER_CUBIC` — bicubic over a 4x4 neighbourhood. Sharper than bilinear when enlarging, slower, and it can overshoot into slight ringing at strong edges. - `cv2.INTER_AREA` — pixel-area resampling, the documented preferred method for image decimation. When enlarging it behaves similarly to nearest-neighbour, so it is not a general-purpose default. - `cv2.INTER_LANCZOS4` — an 8x8 windowed sinc, the sharpest and slowest of the common options for enlargement. ## The practical rules 1. **Shrinking?** `INTER_AREA`. 2. **Enlarging and speed matters?** `INTER_LINEAR`. **Enlarging and quality matters?** `INTER_CUBIC` or `INTER_LANCZOS4`. 3. **Masks, label maps, id images?** `INTER_NEAREST`, always, in both directions. Bilinearly resizing a mask whose values are class ids 0, 1, 7 produces values like 3.5 that correspond to no class at all, and after rounding you get phantom regions of a class that was never there. 4. **Feeding a neural network?** Match whatever the training preprocessing used, including the library. Resize implementations differ in filter support and pixel-centre conventions, and a mismatch is a real, measurable accuracy loss that is invisible in the code. ## The dsize gotcha `dsize` is `(width, height)` — the opposite order of `img.shape`, which is `(rows, cols)` = `(height, width)` for a grayscale image and `(rows, cols, channels)` for colour. Writing `cv2.resize(img, img2.shape[:2])` silently transposes the target dimensions unless the target happens to be square. The safe idioms are `cv2.resize(img, (w, h))` with named variables, or `cv2.resize(img, dsize=(0, 0), fx=0.5, fy=0.5)` to scale by factors and let OpenCV compute the size. Note that when you use `fx`/`fy`, `dsize` must be `(0, 0)` or `None`. ## Extreme reductions Even `INTER_AREA` in a single step is not always the best answer for very large factors on very noisy input; a classic alternative is repeated halving with `cv2.pyrDown`, which applies a Gaussian before each 2x decimation, giving a smooth multi-step reduction. In practice `INTER_AREA` is close enough for a single large step and is far simpler. ## The interview signal The weak answer is "INTER_CUBIC, because it is the highest quality". Quality is direction-dependent: bicubic on a 10x reduction is *worse* than area averaging because it still ignores most of the source. The strong answer names the direction first, explains aliasing versus interpolation, and adds the mask exception unprompted.
- Why is cv2.INTER_NEAREST the right choice for a segmentation mask?Mask values are class ids, not intensities. Bilinear or area interpolation averages them, so ids 0 and 7 produce an in-between value that names no class, and rounding turns it into a phantom region. INTER_NEAREST copies an existing label, guaranteeing the output values are a subset of the input values.
- What is the argument order of dsize in cv2.resize?(width, height), which is the reverse of img.shape's (rows, cols). Passing a shape tuple directly transposes the result for any non-square image and is a frequent bug. Either build the tuple explicitly from named width and height variables, or pass dsize=(0, 0) with fx and fy scale factors instead.
- Someone downscales with INTER_CUBIC because it is the highest-quality flag. What do you tell them?Quality depends on direction. Bicubic still samples a 4x4 neighbourhood around one point, so at a 10x reduction the vast majority of source pixels contribute nothing and fine detail aliases into moire. INTER_AREA averages the full source region per output pixel, which is both cheaper and visibly better for decimation.
- Does the choice matter when resizing inputs for a trained model?Yes. Match the preprocessing used at training time, including which library did the resize. Implementations differ in filter support and pixel-centre convention, so a model trained on area-downscaled crops and served bilinear-downscaled ones sees a subtly different distribution — a measurable accuracy drop with no error anywhere in the code.
saying these in an interview costs you the question
- Uses INTER_CUBIC for everything because it is highest quality
- Passes img.shape as dsize and transposes the image
- Bilinearly resizes a label mask
- Thinks the default INTER_LINEAR is safe at any scale
- Ignores which resize the model was trained with