An improved Deeplabv3+ semantic segmentation algorithm with multiple loss constraints.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Current semantic segmentation algorithms struggle with low accuracy and imprecise object boundaries.
• Limitations in receptive field range and feature fusion hinder performance.

Purpose of the Study:

• To propose an improved semantic segmentation algorithm addressing accuracy and boundary segmentation issues.
• To enhance feature extraction and model optimization through novel architectural designs and loss functions.

Main Methods:

• A multi-level cascading residual structure was employed to expand the network's receptive field.
• A parallel network architecture was designed for extracting multi-depth features.
• Feature fusion and multiple loss constraints were utilized for network optimization.

Main Results:

• The proposed algorithm demonstrated superior performance on the Cityscapes and CamVid datasets.
• Achieved a mean Intersection over Union (MIoU) improvement of 3.07% and 3.59% over the Deeplabv3+ algorithm.
• Significantly improved segmentation accuracy and object boundary delineation.

Conclusions:

• The proposed semantic segmentation algorithm effectively overcomes limitations of existing methods.
• The combination of multi-level cascading residual structure and multiple loss constraints leads to enhanced segmentation performance.
• This approach offers a promising direction for advancing semantic segmentation in computer vision.