How AI reads an Indian food label, and where it slips
LabelSaaf is an AI-powered food label scanner for India: AI reads the nutrition table and ingredients from a photo of the pack, and published rules score it out of 100 with the reason behind every point.
This page explains the reading half: what the AI looks for on the back of a pack, the places a reading can go wrong, and how to catch it when it does.
What the AI reads from one photo
- The nutrition table: energy, protein, carbohydrate, total and added sugars, total and saturated fat, trans fat, fibre and sodium.
- The basis of each column: per 100 g, per 100 ml or per serving. Scores always use per 100 g, so when a pack prints only per-serving figures they are converted using the serving size printed on the pack.
- The ingredient list, including INS numbers, so additives can be named in plain words.
- The product name, which is often the first thing you see while the rest of the reading finishes.
Reading a photo, rather than looking up a barcode, is what lets a pack nobody has catalogued still get an answer. A barcode app can only answer for packs already in its database.
Where a reading can slip
Curved tins and pouches
On a curved panel the rows bend away from the lens, and a figure can end up read against the label one row above or below it. Photograph the table as flat and square as you can.
A zero that was never printed
The misreads that matter most are zeros: a figure the pack does not print, or prints faintly, read as 0. A wrong zero can make a pack look better than it is, which is why every figure behind a score is shown beside it, so you can check a 0 against the pack in your hand.
Two columns side by side
Indian labels often print per 100 g and per serving next to each other, and sometimes a percentage of daily value as a third column. Reading the wrong column changes every figure at once. The verdict says which basis it used.
Glare, blur and cropping
Shiny wrappers reflect light across the numbers, and a table cut off at the edge loses its last rows. When a photo cannot be read, LabelSaaf asks for another rather than guessing.
How to get a clean read
- Fill the frame with the nutrition table, not the whole pack.
- Tilt the pack away from overhead lights until the glare moves off the numbers.
- Flatten a pouch with your thumb; hold a tin so the table faces you square.
- If a figure on the verdict does not match the pack, tell us with Did we read this pack right? It changes what we fix.
Common questions
Does AI decide the score?
No. AI reads the label. The score comes from published rules applied to the figures that were read, so the same figures always get the same score. How the score is calculated.
What can AI misread on a label?
Mostly the things above: a row out of line on a curved pack, a zero that was not printed, or the wrong column. The figures are always shown, so a misread is something you can see.
Is my photo kept?
Usually not: the photo is scored and discarded. It is kept privately only when the read fails or looks unreliable, or when you send it with feedback, and then only to fix the read. Privacy policy.
Photograph the nutrition table on a pack and get a score out of 100, with the reason behind every point.
Scan a label