JPEG Image

The Compression Pipeline

JPEG compression runs through five sequential stages. Each stage targets a different perceptual redundancy in human vision.

Color space conversion transforms RGB triplets into YCbCr: one luminance channel (Y) carrying brightness, and two chrominance channels (Cb, Cr) carrying color difference signals. Human eyes resolve spatial detail in brightness roughly four times better than in color, so separating these channels creates an immediate optimization target.

Chroma subsampling exploits that gap. In 4:2:0 mode — the default for most cameras and encoders — each chrominance channel is downsampled by 2x in both dimensions. Four luminance samples share a single Cb and a single Cr value. This alone cuts raw data by 50% with almost no perceptible change in photographs. 4:2:2 subsamples horizontally only, preserving vertical color edges — useful for text overlaid on photos. 4:4:4 keeps full chrominance resolution, used in medical imaging and professional color grading where any chroma loss is unacceptable.

8x8 block splitting divides each channel into non-overlapping blocks of 64 pixels. Every block is processed independently. This is fast but creates the characteristic "blockiness" visible at low quality settings — adjacent blocks have no shared context.

Discrete Cosine Transform converts each 8x8 block from spatial pixel values into 64 frequency coefficients. The DC coefficient (position 0,0) stores average block brightness. The remaining 63 AC coefficients represent increasing spatial frequencies — low frequencies capture gradual tonal shifts, high frequencies capture sharp edges and fine texture. Most photographic content concentrates energy in the low-frequency coefficients.

Quantization is where data is permanently discarded. Each DCT coefficient is divided by a value from a quantization table and rounded to the nearest integer. High-frequency coefficients get divided by larger numbers, often rounding to zero entirely. The encoder then zigzag-scans the 64 values (low frequency to high), run-length encodes the resulting zero runs, and applies Huffman coding to the output. Arithmetic coding is an alternative entropy coder that produces files 5-10% smaller, but it saw limited adoption for decades due to Unisys and IBM patent claims. Modern encoders like MozJPEG support it freely.

Quality Factor and Quantization Tables

The "quality" slider in image editors scales the quantization table. At quality 100, divisors are small and almost no data is lost — but files are 5-8x larger than quality 85. At quality 50, divisors are aggressive and high-frequency detail vanishes. The relationship is nonlinear.

A 12-megapixel photo (4000x3000) at quality 95 produces roughly 8-12 MB. The same image at quality 85 drops to 2-4 MB. At quality 50, it falls to 500 KB-1 MB. JPEG quality 85 and quality 95 look identical on screens below 6 inches but the file size doubles. This makes quality 80-85 the standard choice for web delivery — the perceptual return above that threshold doesn't justify the bandwidth cost.

Progressive vs Baseline Scan Modes

Baseline JPEG encodes blocks top-to-bottom, left-to-right in a single pass. The image renders line by line as bytes arrive. Progressive JPEG encodes multiple scans: the first pass contains only the DC coefficients (a blurry thumbnail of the full image), and subsequent passes add AC coefficients in bands, sharpening the image incrementally. On slow connections, progressive files appear to "fade in" from blurry to sharp rather than painting downward.

Progressive files are typically 2-6% smaller than baseline at the same quality because Huffman tables span the entire image rather than a single scan. MozJPEG defaults to progressive encoding for this reason.

File Structure and Markers

Every JPEG begins with FF D8 (Start of Image) and ends with FF D9 (End of Image). Between those markers sit segments identified by two-byte codes:

• APP0 (FF E0) — JFIF header with version, resolution, and optional thumbnail
• APP1 (FF E1) — EXIF metadata block (camera settings, GPS, timestamps) and XMP data
• DQT (FF DB) — Quantization tables
• SOF0 (FF C0) — Start of Frame (baseline) with image dimensions, component count, subsampling factors
• DHT (FF C4) — Huffman tables
• SOS (FF DA) — Start of Scan, followed by entropy-coded image data

EXIF Metadata and Privacy

Digital cameras and smartphones embed EXIF data in the APP1 segment: camera model, lens, ISO, aperture, shutter speed, focal length, timestamp, orientation flag, and — critically — GPS coordinates accurate to a few meters. Sharing an unstripped JPEG from a phone can expose your exact location.

Strip all metadata before publishing: exiftool -all= -overwrite_original photo.jpg. Social media platforms strip EXIF on upload, but email attachments, direct file shares, and many CMS platforms preserve it.

Artifacts and Degradation

Three artifact types are characteristic of JPEG:

Blocking artifacts appear as visible 8x8 grid patterns in smooth gradients, most obvious at quality settings below 60. Mosquito noise creates shimmering halos around high-contrast edges — dark text on white backgrounds shows this clearly. Color bleeding occurs at sharp chrominance boundaries when chroma subsampling smears color across the edge.

Re-encoding a JPEG applies quantization again. Each save cycle degrades quality further, even at quality 100, because rounding errors accumulate. Edit in a lossless format (TIFF, PNG) and export to JPEG only as the final step. For lossless operations on existing JPEGs — rotation, cropping on 8x8 boundaries, metadata editing — use jpegtran or MozJPEG's jpegtran fork, which manipulate DCT coefficients directly without re-quantization.

When to Use JPEG vs Alternatives

AVIF outperforms JPEG on compression efficiency, but browser support reached full coverage only in 2024. WebP is a strong middle ground — lossy and lossless modes, transparency, animation — with universal browser support since 2023. JPEG remains the safe default when compatibility with legacy systems, email clients, or print workflows matters.

Optimization Tools

MozJPEG re-encodes JPEGs with optimized Huffman tables, trellis quantization, and progressive scanning. Typical savings: 5-15% file size reduction at the same visual quality. jpegtran performs lossless transformations — rotation, crop, progressive conversion — without touching pixel data. Guetzli (Google) produces files 20-30% smaller than libjpeg but encoding is extremely slow (minutes per image). For batch web optimization, MozJPEG with quality 80-85 and progressive mode hits the best balance of speed, size, and quality.