Lightweight DeepLabv3+ for Semantic Food Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual food analysis systems are crucial for well-being, with food segmentation being a key component.
• Current deep learning methods for food segmentation are computationally intensive, limiting their use on low-performance devices.

Purpose of the Study:

• To propose a novel, lightweight deep learning method for semantic food segmentation.
• To enable efficient and cost-effective food image analysis on resource-constrained devices.

Main Methods:

• Adapted the DeepLabv3+ model by optimizing the backbone with EfficientNet-B1.
• Replaced Atrous Spatial Pyramid Pooling (ASPP) with Cascade Waterfall ASPP (CWASPP).
• Refined encoder output using squeeze-and-excitation attention mechanism and validated on multiple datasets.

Main Results:

• Achieved high performance in semantic food segmentation with significantly lower computational costs.
• The proposed method's results are comparable or superior to state-of-the-art techniques.
• Demonstrated effectiveness on both public and a newly introduced self-acquired food image dataset.

Conclusions:

• The developed lightweight deep learning model is suitable for food image segmentation on low-performance devices.
• This research paves the way for more efficient, cost-effective, and scalable food analysis applications.
• Highlights the potential of optimized deep learning for real-world applications with limited computational resources.