Open-Vocabulary Object Detection Neural Network

The patent implements open-vocabulary object detection where the system can detect objects in any category by processing text queries through a text encoder that generates embeddings. The classification network compares object embeddings with query embeddings to determine object categories, allowing the same system to handle diverse object types without requiring category-specific detection modules. This universal approach enables the system to adapt to any object category while maintaining a unified detection architecture.

The patent segments the processing architecture into independent components: a text encoder that processes text queries separately, an image encoder that processes images to generate object embeddings, and a classification network that combines these representations. This segmentation allows text and image queries to be processed independently through their respective encoders, improving inference efficiency by parallelizing operations. The classification network then integrates these separate processing results to produce final object detection outputs.

The patent applies contrastive pre-training as a preliminary action before the main object detection task. During pre-training, the text encoder and image encoder are trained to generate embeddings that capture semantic relationships between text and image representations. This preliminary training establishes robust feature representations and alignment between modalities, which then improve detection accuracy during the main task. By performing this complex pre-training operation beforehand, the system achieves enhanced performance while separating the computationally intensive pre-training phase from the inference phase.