Abstract
Advancements in artificial intelligence, particularly in computer vision, have driven the research and development of visual food analysis systems focused primarily on enhancing people's well-being. Food analysis can be performed at various levels of granularity, with food segmentation being a major component of numerous real-world applications. Deep learning-based methodologies have demonstrated promising results in food segmentation; however, many of these approaches demand high computational resources, making them impractical for low-performance devices. In this research, a novel, lightweight, deep learning-based method for semantic food segmentation is proposed. To achieve this, the state-of-the-art DeepLabv3+ model was adapted by optimizing the backbone with the lightweight network EfficientNet-B1, replacing the Atrous Spatial Pyramid Pooling (ASPP) in the neck with Cascade Waterfall ASPP (CWASPP), and refining the encoder output using the squeeze-and-excitation attention mechanism. To validate the method, four publicly available food datasets were selected. Additionally, a new food segmentation dataset consisting of self-acquired food images was introduced and included in the validation. The results demonstrate that high performance can be achieved at a significantly lower cost. The proposed method yields results that are either better than or comparable to those of state-of-the-art techniques while requiring significantly lower computational costs. In conclusion, this research demonstrates the potential of deep learning to perform food image segmentation on low-performance stand-alone devices, paving the way for more efficient, cost-effective, and scalable food analysis applications.