JPEG
Also known as: jpg, jpeg compression
The most common lossy image format — it shrinks photographs dramatically by discarding visual detail the human eye barely notices, trading some quality for much smaller files.
- Primary domain
- Graphics & Media
- Sub-category
- Photograph Manipulation & Image Compression
In simple terms
JPEG is the format behind most photographs you see online. Its trick is lossy compression: it permanently throws away some image detail to make the file far smaller — but it throws away the details your eyes are least likely to notice, so a JPEG can be a fraction of the original size while still looking fine. That trade-off is why a photo that would be tens of megabytes raw becomes a couple hundred kilobytes as a JPEG, making photography on the web practical.
The Visual Map
graph LR
rgb["RGB pixels"] --> ycc["1. RGB → YCbCr\n(split brightness/colour)"]
ycc --> sub["2. Chroma subsampling\n(half-res colour)"]
sub --> dct["3. DCT on 8×8 blocks\n(→ frequency coefficients)"]
dct --> quant["4. Quantization\n(round off high frequencies)\n← the lossy step"]
quant --> ent["5. Entropy coding\n(lossless pack)"]
ent --> file["JPEG file"]
More detail
JPEG (named for the Joint Photographic Experts Group, 1992) compresses by exploiting limits of human vision in roughly five steps: (1) Color transform to a brightness + colour representation (YCbCr), separating luminance from chrominance; (2) Chroma subsampling, storing colour at lower resolution than brightness because eyes are far more sensitive to brightness detail; (3) Discrete Cosine Transform (DCT), breaking the image into 8×8 blocks and expressing each as a sum of frequency patterns; (4) Quantization, aggressively rounding off the high-frequency (fine-detail) components — this is where data is actually discarded, controlled by a “quality” setting; and (5) Entropy coding, losslessly packing what remains.
Two consequences follow. It’s lossy and generational — each re-save discards more detail, so repeatedly editing and re-saving degrades the image (“generation loss”), and at high compression you see artifacts (blocky 8×8 regions and “ringing” around sharp edges). And it’s bad at sharp edges and text — the frequency-based approach suits smooth photographic gradients but smears crisp lines, which is exactly where PNG wins. JPEG has no transparency; successors (JPEG 2000, WebP, AVIF, JPEG XL) compress better, but classic JPEG remains nearly universal, and its DCT-based approach is the same family of techniques used in video codecs.
Under the Hood
The heart of JPEG is the DCT plus quantization. A DCT turns an 8×8 block of pixels into 8×8 frequency coefficients, concentrating most energy in the top-left (low frequencies). Zeroing the small high-frequency coefficients — what quantization does — throws away detail the eye barely sees while leaving the block recognisable:
import math
N = 8
def dct2(block): # naive separable 2D DCT
out = [[0.0]*N for _ in range(N)]
for u in range(N):
for v in range(N):
s = 0.0
for x in range(N):
for y in range(N):
s += block[x][y]*math.cos((2*x+1)*u*math.pi/16)\
*math.cos((2*y+1)*v*math.pi/16)
cu = (1/math.sqrt(2)) if u==0 else 1
cv = (1/math.sqrt(2)) if v==0 else 1
out[u][v] = 0.25*cu*cv*s
return out
# A smooth gradient block (typical photo content)
block = [[x*8 + y*4 for y in range(N)] for x in range(N)]
coeffs = dct2(block)
kept = sum(1 for r in coeffs for c in r if abs(c) > 5)
print(f"DCT coefficients with |value| > 5: {kept} of {N*N}")
print(f"DC (average) term: {coeffs[0][0]:.1f} most energy is here")
Only a handful of coefficients carry real energy — JPEG stores those precisely and rounds the rest toward zero, which is why photographic blocks compress so well.
Engineering Trade-offs
- Quality vs size. The quantization “quality” knob trades file size against artifacts directly — high quality is near-transparent but large; low quality is tiny but blocky.
- Lossy compaction vs sharp detail. DCT excels on smooth gradients but rings around hard edges and text — great for photos, poor for screenshots and line art.
- Generation loss vs convenience. Re-saving is easy but each cycle re-quantizes and degrades; lossless editing needs a different master format.
- Universality vs efficiency. JPEG decodes everywhere instantly, but AVIF/WebP/JPEG XL reach the same quality at far smaller sizes — at the cost of compatibility and encode time.
Real-world examples
- Practically every photo on a website or shared by phone is a JPEG.
- A camera’s “quality” setting trades file size against compression artifacts — both are the JPEG quantization step in action.
- Over-compressed images showing blocky squares and halos around edges are classic JPEG artifacts.
Common misconceptions
- “JPEG is lossless if I set quality to 100.” Even at maximum quality, JPEG’s chroma subsampling and DCT rounding usually discard some data; for truly lossless images use PNG.
- “Re-saving a JPEG doesn’t hurt it.” Each save re-compresses and loses more detail; repeated edit-and-save cycles visibly degrade the image.
Try it yourself
Watch energy compaction — run a DCT on a smooth block and see almost all the signal land in a few low-frequency coefficients (python3 only):
python3 - <<'EOF'
import math
N=8
def dct2(b):
o=[[0.0]*N for _ in range(N)]
for u in range(N):
for v in range(N):
s=sum(b[x][y]*math.cos((2*x+1)*u*math.pi/16)*math.cos((2*y+1)*v*math.pi/16)
for x in range(N) for y in range(N))
cu=1/math.sqrt(2) if u==0 else 1; cv=1/math.sqrt(2) if v==0 else 1
o[u][v]=0.25*cu*cv*s
return o
block=[[100+x*5+y*3 for y in range(N)] for x in range(N)]
c=dct2(block)
big=[(u,v) for u in range(N) for v in range(N) if abs(c[u][v])>3]
print(f"{len(big)} of {N*N} coefficients matter; rest round to ~0 (the lossy step)")
EOF
Learn next
- Image format — where JPEG sits among the format choices
- PNG — JPEG’s lossless counterpart for sharp graphics and text
- Video codec — extends JPEG’s DCT idea across frames over time
- Fourier transform — the frequency-domain mathematics the DCT belongs to
Read this in a learning path
All paths →This topic is part of a learning path. Start in context to keep prev/next and progress tracking.
Relationships
- Requires
Neighborhood
A visual companion to the relationships above. Click any node to visit that topic.