DNN Training Recognition Device Using GAN Simulation

The patent segments the DNN training process into two distinct stages: pre-training using simulation images (where computational resources can be optimized) and re-training using actual site images (where only necessary portions are updated). This segmentation allows the system to achieve high accuracy while reducing overall power consumption by avoiding full re-training from scratch.

The patent performs preliminary pre-training using simulation images before actual deployment. This preliminary action establishes a foundational model that can be efficiently refined later through re-training with actual site images, reducing the computational burden during the final training phase and thereby lowering power consumption.

The patent divides the training process into pre-training with simulation images and selective re-training with actual site images. By segmenting this way, the system only updates necessary portions of the DNN for real image conditions rather than re-training the entire network, thus improving recognition accuracy for real images while reducing training time.

The patent applies local quality by updating only the necessary portions of the DNN that are relevant to actual site image conditions during re-training, while maintaining the pre-trained parameters for other parts of the network. This localized update approach improves recognition accuracy for real images without the time cost of re-training the entire DNN.

The patent introduces actual site images as an intermediary between simulation images and the final DNN model. The process uses simulation images for initial pre-training (reducing manual annotation), then introduces actual site images through re-training to bridge the gap between simulation and real-world conditions, thereby improving recognition accuracy while maintaining the benefits of automated simulation-based training.

The patent performs preliminary pre-training using simulation images with automated annotation, then conducts subsequent re-training using actual site images to refine the model. This preliminary action with simulation images reduces manual annotation efforts, while the follow-up re-training with actual images ensures high recognition accuracy by adapting to real-world conditions.

The patent segments the training process into pre-training and re-training phases, allowing comprehensive learning to be achieved through staged computation rather than requiring full computational resources simultaneously. This segmentation enables the system to reach comprehensive learning capability while reducing peak hardware requirements.

The patent performs preliminary pre-training with simulation images to establish comprehensive learning foundations, then conducts re-training with actual site images to refine and adapt the model. This preliminary action allows comprehensive learning to be built incrementally, reducing the need for extensive hardware resources during the final training phase.