Media Analysis

The trend and behaviour analysis performed by the iCOP engine identifies candidate child sexual abuse (CSA) media. This media is then sent to the Media Analysis component to be verified for pornographic content and prioritised. Cutting-edge feature extraction and machine learning technology are used to:

classify images or videos as harmless or potentially child pornographic (content classification)

perform a similarity matching with known CSA content to indicate if the discovered content is entirely new or already known (similarity search).

Content Classification
iCOP's content classification engine performs an automatic categorisation of pictures and video scenes into "harmless", "offensive", or "CSA". Though such visual recognition is a challenging problem (and detection rates are far from human accuracy), it offers an efficient prioritisation or filtering of evidence even in cases where queries or filenames (like "000256.jpg") are not indicative. The iCOP approach offers two key benefits:

Multimodality: iCOP analyzes the colour and texture of images. Beyond this, for videos it also exploits other modalities like motion information and the audio stream. By combining these different clues in a joint optimisation, the accuracy of content classification is improved significantly.

Adaptation to CSA: Content classification technology often focuses on a detection of general pornographic content, which is not suitable for detecting child pornography. Therefore, in iCOP we collaborate with European police partners to benchmark and optimise our approaches specifically for the detection of CSA content.

Similarity Search
When CSA material is detected, it is important for investigators to know whether it is entirely new or already known. Therefore, iCOP supports an efficient matching of candidate material with police databases collected in previous cases. In contrast to standard methods based on file hashes, the iCOP approach supports a similarity-based matching, which allows the identification of near-duplicates (like resized or re-econded images) as well as other content from the same scene or shoot. iCOP supports both images and videos, and exploits techniques of approximate nearest neighbor search to operate on large-scale datasets.