Hand Key Point Recognition via Cycle-GAN Emulation

The patent introduces Cycle-GAN as an intermediary tool to translate virtual images into emulation images that bridge the gap between virtual and real image domains. The Cycle-GAN model learns the mapping between virtual and real image styles, enabling virtual images to be transformed into realistic-looking training samples while preserving the original key-point coordinates, thus resolving the contradiction between sample availability and recognition accuracy

The patent changes the visual parameters of virtual images by applying style transfer through Cycle-GAN. The model modifies color distribution, texture characteristics, and lighting parameters of virtual images to match real image properties, while maintaining the semantic content and key-point information. This parameter transformation enables virtual images to resemble real images closely enough for accurate gesture recognition training

The patent uses virtual images as copies of real images, which automatically contain accurate key-point coordinates through the three-dimensional modeling process. Instead of manually labeling real images, the system generates synthetic virtual images that replicate the structural and geometric properties of real hands, complete with precise key-point annotations already embedded in the generation process, thereby eliminating manual labeling while maintaining data accuracy

The patent performs preliminary action by pre-computing and embedding accurate key-point coordinates during the virtual image generation phase through three-dimensional modeling. This preliminary labeling action eliminates the need for subsequent manual annotation, as the key-point data is automatically available and accurate before the training process begins