Image Resizing & Resampling Algorithms: Nearest-Neighbor, Bilinear, Bicubic, and Lanczos

The Mathematics of Spatial Image Resampling

In digital image processing, a bitmap image is a two-dimensional grid of discrete spatial samples: $$I(x, y) = [R, G, B, A]^T \quad \text{where } x \in [0, W-1], y \in [0, H-1]$$

When you resize an image from source dimensions $(W_s, H_s)$ to target dimensions $(W_t, H_t)$, target coordinates rarely align with exact source pixel centers. The scaling ratio is defined as: $$s_x = \frac{W_s}{W_t}, \quad s_y = \frac{H_s}{H_t}$$

For any fractional continuous target coordinate $(u, v)$, the image engine must interpolate (reconstruct) the missing continuous signal from surrounding discrete samples before re-sampling onto the target grid.

Comparison of Core Resampling Filters

Filter Complexity & Visual Quality Trade-off:

Nearest-Neighbor ───> Bilinear ───> Bicubic (Catmull-Rom) ───> Lanczos-3
(1 sample, sharp/pixelated)   (4 samples, soft)       (16 samples, smooth curves)   (36 samples, high acutance)

1. Nearest-Neighbor Interpolation

The simplest filter assigns each destination pixel the exact value of the closest source pixel: $$f(x) = \text{rect}(x) = \begin{cases} 1 & \text{if } |x| < 0.5 \ 0 & \text{otherwise} \end{cases}$$

Computational Cost: $O(1)$ operations per pixel.
Strengths: Zero blurring; preserves sharp color boundaries. Indispensable for pixel art, icon palettes, and QR codes.
Weaknesses: Severe aliasing, jagged diagonal edges, and staircase distortion when enlarging photographic content.

2. Bilinear Interpolation

Computes a weighted average of the 4 nearest surrounding source pixels $(2 \times 2$ grid) using linear distance weighting: $$f(x, y) = (1-s)(1-t)I_{0,0} + s(1-t)I_{1,0} + (1-s)tI_{0,1} + stI_{1,1}$$

Computational Cost: $O(4)$ operations per pixel.
Strengths: Fast execution supported in hardware by GPU texture mapping units (TMUs).
Weaknesses: Attenuates high spatial frequencies, causing soft blurring and slight contrast loss along edges.

3. Bicubic Interpolation (Catmull-Rom Spline)

Samples 16 neighboring pixels ($4 \times 4$ neighborhood) weighted by a cubic polynomial spline: $$W(x) = \begin{cases} (a+2)|x|^3 - (a+3)|x|^2 + 1 & \text{for } |x| \le 1 \ a|x|^3 - 5a|x|^2 + 8a|x| - 4a & \text{for } 1 < |x| < 2 \ 0 & \text{otherwise} \end{cases}$$ Setting parameter $a = -0.5$ yields the Catmull-Rom spline, which balances smooth tonal gradients with edge preservation.

Computational Cost: $O(16)$ operations per pixel.
Strengths: Significantly reduces blockiness while retaining crisp curves; the gold standard for portrait and landscape photography.

4. Lanczos Resampling (Sinc-Windowed Reconstruction)

Approximates the mathematically optimal Whittaker-Shannon sinc reconstruction filter windowed over an $N$-lobe radius: $$L(x) = \begin{cases} \text{sinc}(x) \cdot \text{sinc}(x/a) & \text{for } -a < x < a \ 0 & \text{otherwise} \end{cases}$$

Computational Cost: Up to $O((2a)^2)$ operations (36 samples for Lanczos-3).
Strengths: Maximum acutance and edge sharpness; ideal for photographic downsampling and print prepress.

Summary Matrix: Selecting the Right Algorithm

Resampling Algorithm	Kernels Sampled	Primary Use Case	Edge Acutance	Computational Load
Nearest-Neighbor	1 pixel	Pixel art, retro graphics, QR codes	Very High (Jagged)	Negligible ($O(1)$)
Bilinear	4 pixels	Fast mobile previews, real-time video	Moderate (Soft)	Low ($O(4)$)
Bicubic	16 pixels	General photography, web banners	High (Smooth)	Moderate ($O(16)$)
Lanczos-3	36 pixels	Downsampling high-res camera RAWs	Maximum (Crisp)	Intensive ($O(36)$)

In Mega Toolskit's Image Tools, our client-side canvas engine automatically deploys multi-step mipmapped downsampling to avoid moiré patterns and maintain high visual fidelity.