Our #image #analysis #pipeline is ready to go into #pruduction.
It combines #AI and #forensic techniques to assess authenticity. It first extracts EXIF/XMP #metadata to detect real camera data forcing a “photograph” classification if present. Then, a CLIP based preprocessor (#computervision) classifies images into categories: photograph, meme, infographic, comic/cartoons, or text only.
Custom configurations (#development) adjust weights so that, for example, memes/text only images discount AI model influence while forensic tests (ELA, FFT, noise, edge sharpness, text detection) (#science, #math) quantify artifacts. #C2PA data is also considered. All evidence is combined to yield a final decision which will be the “confidence percentage” data that will be included in every post moving forward.
