Neural Network Sub-Type Recognition for Rare Object Detection

The training process is segmented into two distinct phases: a first training phase using a first neural network module with available training images, and a second training phase using a second neural network module with synthesized training images. This segmentation allows the system to handle rare object types by separating the training into manageable stages, first learning from real data and then reinforcing with synthesized data, thereby improving recognition accuracy without requiring a large quantity of real training images for rare object types.

The system performs preliminary action by first training the first neural network module with available real training images before synthesizing and using virtual training images. This preliminary training establishes a foundation with real data, and subsequent training with synthesized images reinforces and extends this foundation. This approach allows the system to achieve sufficient training for rare object types without requiring extensive time to evaluate and collect additional real training images, as the synthesized images supplement and extend the training data efficiently.

The system employs self-service by using the first neural network module to automatically extract and generate training images for the second neural network module without requiring manual intervention from experts. The first neural network module processes real training images and generates synthesized training images that are then used for further training. This self-service mechanism eliminates the need for experts to manually determine and annotate specific image features for rare object types, thereby maintaining high classification accuracy while significantly reducing training complexity and manual outlay.