FE-Net: Feature enhancement segmentation network.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Semantic segmentation is crucial for image analysis, aiding tasks like measurement and feature selection.
• Existing methods struggle with precise class edge delineation, particularly in multi-class regions.
• Complex backgrounds and fine object boundaries pose significant challenges for current segmentation models.

Purpose of the Study:

• To introduce a novel semantic segmentation approach, the Feature Enhancement Network (FE-Net).
• To enhance segmentation performance, particularly at class edges and in multi-class scenarios.
• To improve the precision and effectiveness of image segmentation in complex environments.

Main Methods:

• Developed the Feature Enhancement Network (FE-Net) incorporating a Smart Edge Head (SE-Head).
• Implemented a transitional structure combining SE-Head with FCN-Head and SepASPP-Head for gradual loss weight transition.
• Introduced a pixel-wise weight evaluation method, a pixel-wise weight block, and a feature enhancement loss.

Main Results:

• FE-Net demonstrated significant performance improvements on Pascal VOC2012, SBD, and ATR datasets.
• Achieved best mean Intersection over Union (mIoU) enhancements of 15.19%, 1.42%, and 3.51% respectively.
• Showcased superior effectiveness in segmenting key pixels on the Pole&Hole match dataset.

Conclusions:

• FE-Net effectively addresses limitations in semantic segmentation, particularly edge precision in complex scenes.
• The proposed methods, including SE-Head and pixel-wise weighting, enhance segmentation accuracy.
• FE-Net offers a robust solution for detailed image segmentation tasks, outperforming existing methods.