Image Scene Segmentation via Multi-Stage Fusion and Block Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning networks for image scene segmentation struggle with accurate segmentation of images containing multiple categories of scenes, often resulting in fragmented segmentation results that affect downstream services.
Innovation Solution
The method involves obtaining an intermediate scene segmentation image through initial segmentation and fusion, detecting segmentation blocks for processing, and performing segmentation correction to obtain a target scene segmentation image, thereby reducing fragmentation and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning networks are used for image scene segmentation, then segmentation capability is improved, but segmentation accuracy deteriorates for multi-category scenes due to fragmented results
Solution Approach 1:
The patent divides the scene segmentation task into multiple processing stages: initial segmentation to identify candidate regions, fusion to combine segmentation results from different networks, and post-processing to refine boundaries. This multi-stage segmentation approach resolves the contradiction by handling different scene categories at appropriate processing depths, preventing fragmentation while maintaining adaptability across diverse scene types
Solution Approach 2:
The patent combines segmentation results from multiple deep learning networks through a fusion mechanism that aggregates predictions across different architectures. By merging the strengths of various networks rather than relying on a single model, the system achieves both high adaptability to different scene categories and improved segmentation accuracy, eliminating the fragmented results that plague individual networks
2Measurement precision
If more sophisticated deep learning networks are used, then segmentation accuracy is improved, but computing requirements increase making tradeoff difficult
Solution Approach 1:
The patent applies partial action by using multiple specialized networks for different scene categories rather than one excessively complex universal network. Each network is optimized for specific categories, achieving high accuracy where needed while keeping individual network complexities manageable. The fusion mechanism then combines these partial results into a complete segmentation solution
Solution Approach 2:
The patent creates a universal segmentation framework that can handle multiple scene categories through a common fusion architecture. This multi-functional system processes inputs from various specialized networks and produces unified segmentation results, achieving broad accuracy across categories without requiring each component to be excessively complex
3Reliability
If deep learning networks are optimized for scene segmentation, then segmentation results are improved, but fragmentation increases affecting downstream services
Solution Approach 1:
The patent implements feedback mechanisms in the post-processing stage where segmentation results are analyzed for fragmentation, and corrective actions are applied to merge fragmented regions. The system uses boundary refinement and region merging algorithms that respond to detected fragmentation, providing feedback loops that continuously improve result continuity while maintaining segmentation quality
Solution Approach 2:
The patent performs preliminary fusion of segmentation results before final output, combining predictions from multiple networks in advance to prevent fragmentation. By merging results early in the processing pipeline rather than treating them separately, the system establishes continuous segmentation boundaries from the outset, ensuring both quality and stability of the final results
Data Source
AI summary
The embodiments of the present disclosure provide an image scene segmentation and apparatus, and a device and a storage medium. The method includes: obtaining an intermediate scene segmentation image by performing scene initial segmentation and scene initial fusion on an obtained target image; detecting, from the intermediate scene segmentation image, segmentation blocks to be processed; and obtaining a target scene segmentation image of the target image by performing segmentation correction on the segmentation blocks to be processed.