A multi-modal approach to detect inappropriate cartoon video contents using deep learning networks

Malicious or Benign? Towards Effective Content Moderation for Children's Videos

“…Many solutions have been proposed to the problem of detecting video content that is inappropriate for children. A general dichotomy between these techniques is whether they rely on language classification models and leverage video meta-data, user comments, and sub-titles as inputs (Gkolemi et al 2022;Papadamou et al 2021;Alshamrani et al 2021), use the video stream alone (Yousaf and Nawaz 2022;Singh et al 2019;Ishikawa, Bollis, and Avila 2019), or use a multimodal approach that combines both vision and language cues (Chuttur and Nazurally 2022;Le, Tandon, and Oinar 2022;Tahir et al 2019;Kaushal et al 2016).…”

“…Only a small set of prior work has focused on cartoon videos (Chuttur and Nazurally 2022;Kaushal et al 2016;Ishikawa, Bollis, and Avila 2019), whereas most other authors have studied videos with human characters, movie clips, news, and sports. Elsagate (Placido 2017), the misuse of popular cartoon characters to create clickbait content, triggered a surge of interest in content moderation for cartoons.…”

“…Since less attention has been devoted to cartoons, we focus on cartoon videos only and have collected a dataset of malicious cartoon videos that are difficult to detect. Chuttur and Nazurally (2022) proposed a multimodal solution using text (user comments and captions), along with image data. As images do not capture the motion related features it is likely that malicious spatio-temporal patterns will be ignored by their method.…”

“…A variety of approaches have been used to tackle this problem including leveraging user comments (Alshamrani et al 2021), metadata (Gkolemi et al 2022;Papadamou et al 2021) and video features (Yousaf and Nawaz 2022;Singh et al 2019;Ishikawa, Bollis, and Avila 2019); others have employed multimodal techniques rather than relying on a single data source (Chuttur and Nazurally 2022;Papadamou et al 2021;Singh et al 2019;Kaushal et al 2016). We focus on analyzing and annotating subtle audiovisual features present in cartoon videos which are not conducive for the overall mental and physical development of toddlers and pre-school kids.…”