Zicheng Gao, Yonghua Huang, Xufeng Yuan, Hao Guo, Francesco Marinello, Lorenzo Guerrini, Alberto Carraro
Abstract
Reducing food waste is crucial to advancing sustainable development goals and creating a more sustainable food system. However, traditional Waste Manag. methods, which rely on manual weighing and recording, are inefficient. To address this challenge, this study employs computer vision to improve the understanding of food waste and proposes an effective multimodal learning approach for the semantic segmentation of food waste. We created a specialized food waste image dataset. It comprises red-green-blue-depth (RGB-D) images collected from college canteens. First, we extracted key frames of food waste from the raw collected images. Subsequently, we proposed a novel multimodal semantic segmentation network, which introduces a “feature rectification and fusion” module to enhance information interaction and fusion across different data modalities. Finally, the segment anything model is incorporated to refine semantic segmentation outputs. Ablation experiments demonstrated the effectiveness of the proposed module. Comparative experiments show that our method outperformed existing approaches on our dataset, achieving a “mean intersection over union” score of 71.55 % and a “mean pixel accuracy” score of 83.44 %.
Keywords
Food waste recognition; Semantic segmentation; RGB-D image; Multimodal; College canteen
1-s2.0-S0921344925005452-main.pdf
