Training Image Data Generation via Object Removal

The patent applies preliminary action by pre-segmenting images into multiple patches before training the object recognition model. This preprocessing step creates a structured foundation that enables automated label generation later, reducing the need for manual labeling while maintaining accuracy. The patches are created in advance with their spatial relationships preserved, allowing the system to efficiently generate training data through automated processes.

The patent uses copying by generating synthetic training images through template matching and patch reconstruction. Instead of manually labeling each image, the system creates copies and variations of annotated patches, combining them to form new training images with automatically generated labels. This copying approach maintains the realism and accuracy of manual labeling while dramatically reducing the time required.

The patent applies local quality by ensuring that each patch retains its original local characteristics and visual fidelity from the source images. Instead of generating entirely synthetic images that may lack realism, the system preserves the authentic local properties of image patches while arranging them in synthetic configurations. This approach maintains the visual realism needed for robust network training while achieving automated generation.

The patent uses composite materials by combining multiple real image patches with template structures to create composite training images. These composite images retain the authentic visual properties of real images while incorporating the structured organization needed for automated generation. The resulting training data achieves both productivity and reliability by merging real image content with synthetic arrangement.

The patent applies segmentation by dividing images into discrete patches that can be independently processed and recombined. This segmentation approach enables the system to generate multiple training samples from single images by rearranging patches, effectively increasing training data volume without requiring complex transformations that would misalign labels. Each patch maintains its original label information, ensuring accurate label alignment.

The patent uses another dimension by transitioning from image-level transformations to patch-level recombination. Instead of applying transformations that affect the entire image and risk label misalignment, the system operates at the patch level, where each patch's label remains independently valid. This dimensional shift from global to local operations preserves label accuracy while generating diverse training samples.