Multibranch semantic image segmentation model based on edge optimization and category perception.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Current semantic image segmentation methods often overlook multiscale feature interactions, leading to information loss and inaccurate object boundary segmentation.
• Deepening convolutional layers exacerbates the loss of spatial details, negatively impacting segmentation accuracy, especially at object edges.

Purpose of the Study:

• To propose an edge optimization and category-aware multibranch semantic segmentation network (ECMNet) to address limitations in current segmentation methods.
• To improve the accuracy of semantic image segmentation, particularly at object boundaries, and reduce category confusion.

Main Methods:

• An attention-guided multibranch fusion backbone connects features of different resolutions in parallel for multiscale information interaction.
• A category perception module learns category representations and uses attention to guide pixel classification for accuracy.
• An edge optimization module integrates edge features adaptively to enhance edge expression and segmentation.

Main Results:

• ECMNet achieved a mean Intersection over Union (MIoU) of 79.2% on the Cityscapes dataset and 79.6% on the CamVid dataset.
• The proposed network demonstrates a significantly lower parameter count compared to existing models.
• The method effectively improves semantic image segmentation performance and resolves partial category segmentation confusion.

Conclusions:

• ECMNet offers a promising approach for semantic image segmentation, achieving high accuracy and efficiency.
• The network's ability to handle multiscale features and optimize edge details provides a robust solution for complex segmentation tasks.
• ECMNet shows significant application potential in various computer vision domains requiring precise image segmentation.