Semantic Segmentation via Multiscale Feature Merging and Confidence Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition systems face challenges in detecting small, cluttered, or occluded objects in images, often incorrectly identifying parts of objects as whole objects or combining different object portions into a single entity.
Innovation Solution
The method involves extracting core instance features, calculating feature maps at multiscale resolutions, generating detection boxes, calculating segmentation masks, merging masks, and refining confidence scores using auxiliary networks to improve object detection and segmentation, particularly in crowded scenes and for small objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based systems are used to improve object detection accuracy, then detection precision improves, but the system fails to correctly detect small, cluttered, or occluded objects
Solution Approach 1:
The patent applies segmentation by dividing the detection task into multiple stages: generating region proposals, creating segmentation masks for each proposal, and refining detections through non-maximum suppression. This multi-stage segmentation approach allows the system to handle small and occluded objects more effectively by processing potential object regions individually rather than attempting global detection
Solution Approach 2:
The patent introduces a new dimension to object detection by incorporating segmentation masks as an additional output layer. Instead of only predicting bounding boxes, the system generates pixel-level masks that provide fine-grained object boundaries, enabling more accurate detection of small and occluded objects by capturing their precise spatial extent
2Productivity
If the system detects all potential objects, then detection completeness improves, but spurious detections increase
Solution Approach 1:
The patent implements feedback through non-maximum suppression (NMS), where detected objects with high confidence scores provide feedback to suppress overlapping detections. The system uses confidence thresholds and overlap metrics to iteratively filter spurious detections while preserving true objects, balancing completeness and precision
Solution Approach 2:
The patent applies parameter changes by adjusting confidence thresholds and overlap tolerances to control the trade-off between detection completeness and accuracy. By dynamically setting these parameters, the system can adapt to different scenarios and reduce spurious detections while maintaining high detection rates
Data Source
AI summary
Detecting objects in an image includes: extracting core instance features from the image; calculating feature maps at multiscale resolutions from the core instance features; calculating detection boxes from the core instance features; calculating segmentation masks for each detection box of the detection boxes at the multiscale resolutions of the feature maps; merging the segmentation masks at the multiscale resolutions to generate an instance mask for each object detected in the image; refining the confidence scores of the merged segmentation masks by auxiliary networks calculating pixel level metrics; and outputting the instance masks as the detected objects.