A First Approach to Image Transformation Sequence Retrieval

Detecting the corresponding editions from just a pair of input-output images represents an interesting task for artificial intelligence. If the possible image transformations are known, the task can be easily solved by enumeration with brute force, yet this becomes an unfeasible solution for long sequences. There are several state-of-the-art approaches, mostly in the field of image forensics, which aim to detect those transformations; however, all related research is focused on detecting single transformations instead of a sequence of them. In this work, we present the Image Transformation Sequence Retrieval (ITSR) problem and describe a first attempt to solve it by considering existing technology. Our results demonstrate the huge difficulty of obtaining a good performance-being even worse than a random guess in some cases and the necessity of developing specific solutions for ITSR.