Multiscale scene parsing network.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Existing lightweight scene parsing networks struggle to balance multiscale feature representation precision and computational efficiency.
• Efficient extraction of multi-scale information is crucial for accurate semantic segmentation.

Purpose of the Study:

• To propose MSPNet, a novel lightweight multiscale scene parsing network that addresses the limitations of current models.
• To enhance the balance between feature representation accuracy and computational efficiency in lightweight networks.

Main Methods:

• MSPNet utilizes the StarNet backbone for efficient feature transformation.
• The Efficient Pixel Localization Attention (EPLA) module, comprising ELA and PagFM submodules, is innovatively embedded into the PSPNet architecture.
• Depthwise separable convolutions and channel reparameterization techniques are incorporated for model compactness.

Main Results:

• MSPNet achieved a mean Intersection over Union (mIoU) of 87.19% on the Pascal VOC2012 validation set, a 1.79% improvement over PSPNet.
• The model demonstrates comparable GFLOPs and parameter counts to the MobileNet series.
• MSPNet outperforms contemporary lightweight state-of-the-art models in both accuracy and efficiency.

Conclusions:

• MSPNet offers an effective solution for real-time semantic segmentation on resource-constrained mobile devices.
• The proposed EPLA module enhances pixel-level feature localization and cross-scale feature integration.
• MSPNet represents a significant advancement in lightweight scene parsing networks.