An Interactive Image Segmentation Method Based on Multi-Level Semantic Fusion.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Interactive object segmentation is crucial for applications like medical diagnosis and image editing.
• Existing methods struggle to effectively utilize user annotation information for accurate segmentation.
• Analyzing 2D/3D sensor data requires robust object segmentation techniques.

Purpose of the Study:

• To propose a novel interactive image segmentation technique for static images using multi-level semantic fusion.
• To enhance the utilization of user-guidance information for improved segmentation accuracy.
• To develop a method applicable to both 2D and 3D sensor data.

Main Methods:

• Introduced a cross-stage feature aggregation module to prevent semantic information loss during network processing.
• Incorporated a feature channel attention mechanism to refine segmentation edges by capturing richer feature details.
• Utilized user-guidance information both inside and outside the target object for segmentation.

Main Results:

• Achieved approximately 2.1% higher intersection over union (IOU) accuracy compared to existing methods on the PASCAL VOC 2012 dataset.
• Demonstrated improved performance and effectiveness in segmenting objects from static images.
• Showcased finer segmentation edges due to the feature channel attention mechanism.

Conclusions:

• The proposed multi-level semantic fusion method significantly advances interactive image segmentation.
• The technique offers improved accuracy and finer segmentation details, applicable to 2D and 3D data.
• Potential applications in medical imaging, robotics, and integration with other visual semantic analysis workflows.