Forensic Deep Learning: Kaggle Camera Model Identification ChallengeVladimir IglovikovBlockedUnblockFollowFollowingDec 5From https://habr.com/company/ods/blog/415571/There was a computer vision challenge that was hosted at kaggle.com about a year ago named IEEE’s Signal Processing Society — Camera Model Identification..The task was to assign what type of the camera was used to capture an image..The work here is based mainly on this further investigation.Among the many applications for image processing is that of camera model detection..In the simplest case, the camera model should be stored in the image metadata, making camera identification a very easy problem..Different algorithms are used for different camera models, which means that each of them creates model specific artifacts that one may use as features for a machine learning pipeline.And of course, there is plenty of literature regarding the topic..For example, the first idea that came in my mind when I faced this problem was to take an image, subtract smoothed version of it, calculate different statistics of this difference, like mean, median, std, different quantiles and train xgboost on them..One of the first papers on how to do this camera model detection was very close to what I just described..Let’s say we want to do this camera detection for a new, just released model..But such a unified comparison is better than nothing.For the camera detection task, IEEE Signal Processing Society organized a challenge in which 582 teams had two months and real-time leaderboard to compare their methods..Others, like myself, have experience with computer vision techniques but were not even aware that such a problem as camera identification even existed and someone is interested in solving it.The organizers prepared a train set consisting of 2750 images, from 10 cameras..275 per camera.Camera models that we needed to distinguish were:Sony NEX-7Motorola Moto XMotorola Nexus 6Motorola DROID MAXXLG Nexus 5xApple iPhone 6Apple iPhone 4sHTC One M7Samsung Galaxy S4Samsung Galaxy Note 3To make the problem less boring, the organizers used 20 different phones in total to collect the images..For example, if the images in the train data for the iPhone 6 were taken with Ben Hamner’s device (Camera 1), the images in the test data were taken with Ben Hamner’s second device (Camera 2), since he lost the first device in the Bay while kite-surfing.Moreover..Images in the train were in full size, while in test only central 512×512 crops were used.. More details