Stability and Interpretability of Computer Vision Models on the Example of Automatic Content Browsing System

Speakers will talk about the system of automatic content screening, created to find moments that need to be hidden from users due to the requirements of Roskomnadzor. This system is capable of recognizing scenes with smoking, drinking alcoholic beverages, erotic scenes and others.

This system is based on artificial intelligence models that solve the problems of classification, detection of objects. However, traditional methods of training artificial intelligence models have some disadvantages, including those related to the instability of the obtained models even with a slight change in the distribution of input data. In this regard, the obtained models were investigated for stability to such deviations of input data, and robust learning methods were used to improve stability.