Logic Gate Networks Enable Efficient Video Copy Detection

A novel framework utilizing differentiable Logic Gate Networks (LGNs) substitutes traditional floating-point feature extractors with compact, logic-oriented representations for detecting video copies. This method integrates aggressive frame reduction, binary preprocessing, and a trainable LGN embedding model that acquires knowledge of both logical operations and their connections. Once trained, the model is converted into a purely Boolean circuit, allowing for rapid and memory-efficient inference. The research rigorously assesses various similarity techniques, binarization methods, and LGN designs across multiple dataset folds and levels of difficulty. This study tackles the issues of computational expense and descriptor size that deep neural networks face, which restrict their practical use in high-throughput environments. The arXiv paper (2604.21694) showcases this approach as an effective solution for large-scale video copy detection amidst various visual distortions.