Development of a compressed FCN architecture for semantic segmentation using Particle Swarm Optimization.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Conventional deep learning classification networks are adapted for semantic segmentation, creating Fully Convolutional Networks (FCNs).
• These FCN models are computationally expensive, requiring substantial storage and inference time, limiting their use on edge devices.

Purpose of the Study:

• To develop a compressed version of a VGG16-based Fully Convolutional Network (FCN) for efficient semantic segmentation.
• To enable the deployment of accurate FCN models on resource-constrained edge devices.

Main Methods:

• A VGG16-based Fully Convolutional Network (FCN) was compressed using Particle Swarm Optimization.
• The compressed model's performance was evaluated on diverse datasets, including plant disease images, street scenes, and X-ray images.

Main Results:

• The developed compressed FCN model achieved significant savings in storage space and faster inference times.
• The model demonstrated comparable accuracy to standard FCNs, even after an 851x compression ratio.
• The compressed FCN proved implementable on edge devices.

Conclusions:

• Particle Swarm Optimization effectively compresses FCNs for semantic segmentation.
• The compressed FCN offers a viable solution for deploying accurate image segmentation on edge devices.
• This approach balances model efficiency with high performance across various imaging applications.