Preliminary Study on Detection of Breasts

In the realm of digital image manipulation, deep fakes, predominantly sourced from pornographic materials, present a significant challenge, especially prevalent in the form of face and body swapping techniques. This emerging issue involves substituting the faces or bodies of individuals in explicit content, using advanced methods like those demonstrated in the DeepNude application. In response, we present an approach which solves the first part of the pipeline – detection of breasts. There has been limited research regarding this biometric modality, with notable exceptions such as breast cancer identification. Due to the lack of research and the absence of open, freely available data, we developed our own dataset. Images with annotations were acquired from pornhub.com and curated by experts. Annotations include name, cup size, possible breast augmentations, ethnicity, among others. To demonstrate that dataset is challenging enough for future research we used images to train in class-agnostic way three CNN-based detection models. The results show the feasibility of not only the proposed detection approaches for the task, but also the dataset and hopefully pave the way for future applications such as supporting court decisions, enhancing virtual clothing fitting techniques, and more.