Image Object Segmentation Using Reference Patches

The patent copies segmentation masks from reference images that are similar to the target image. By finding images with similar content and copying their pre-computed masks, the system achieves accurate segmentation without requiring manual selection, thus resolving the contradiction between accuracy and time consumption.

The patent performs preliminary segmentation by pre-computing masks for reference images in advance. These pre-computed masks are then reused for target images through similarity matching, eliminating the need for repeated manual or computational segmentation efforts and significantly reducing time consumption.

Instead of relying on saliency detection that fails for cluttered images, the patent copies masks from reference images that are similar to the target image. This approach maintains processing speed while achieving reliable segmentation results even in cluttered scenarios where saliency methods fail.

The patent introduces reference images as an intermediary between the target image and the segmentation result. By using similar reference images with known correct masks, the system bridges the gap between fast automated processing and reliable segmentation results for difficult images.

The patent creates a universal segmentation system that works across all object categories by using content-based similarity matching with reference images. Instead of requiring separate category-specific models, a single system can handle any object type as long as similar reference images are available, thus achieving both accuracy and versatility.

The patent changes the approach from fixed category-specific models to a flexible similarity-based method. By changing the segmentation parameter from category-dependent to content-dependent, the system achieves accuracy for each specific case while maintaining versatility across all categories.