Publicação
Comparison of Image Processing and Classification Methods for a Better Diet Decision-Making
| Resumo: | This paper aims to explore the use of different deep learning techniques, specifically convolutional neural networks (CNNs), for dietary assessment through image food recognition and compare their performance to the human visual system (HVS). Currently, there are three main techniques for using CNNs in this task: training a network from scratch; using an off-the-shelf pre-trained network; and performing unsupervised pre-training with supervised adjustments. In this study, the authors evaluate the performance of three CNN models with varying numbers of parameters (5,000 to 160 million) based on dataset size and spatial image context. The authors also consider human knowledge and classification to compare the performance of the CNNs to the HVS. They find that while the CNNs make errors across different food classes, the HVS tends to make semantic errors with specific food classes. As a result, the HVS shows more consistency in its answers. Overall, the findings suggest that the HVS is more accurate when the dataset is diverse, while the CNN performs better when the dataset is focused on a particular niche. In conclusion, this study provides empirical evidence that machine learning can be more efficient than the HVS in certain tasks but also highlights the strengths and limitations of both approaches. The authors suggest that combining CNNs with other classification techniques, such as bagof-words, may be a promising approach for improving the accuracy of dietary assessment through image food recognition. |
|---|---|
| Autores principais: | Abbasi, Maryam |
| Outros Autores: | Cardoso, Filipe; Martins, Pedro |
| Assunto: | Compare Image processing Machine learning Image Classification GoogLeNet Inception-v3 ResNet Bag-of-words |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | capítulo de livro |
| Tipo de acesso: | acesso restrito |
| Instituição associada: | Instituto Politécnico de Santarém |
| Idioma: | inglês |
| Origem: | Repositório Científico do Instituto Politécnico de Santarém |
| Resumo: | This paper aims to explore the use of different deep learning techniques, specifically convolutional neural networks (CNNs), for dietary assessment through image food recognition and compare their performance to the human visual system (HVS). Currently, there are three main techniques for using CNNs in this task: training a network from scratch; using an off-the-shelf pre-trained network; and performing unsupervised pre-training with supervised adjustments. In this study, the authors evaluate the performance of three CNN models with varying numbers of parameters (5,000 to 160 million) based on dataset size and spatial image context. The authors also consider human knowledge and classification to compare the performance of the CNNs to the HVS. They find that while the CNNs make errors across different food classes, the HVS tends to make semantic errors with specific food classes. As a result, the HVS shows more consistency in its answers. Overall, the findings suggest that the HVS is more accurate when the dataset is diverse, while the CNN performs better when the dataset is focused on a particular niche. In conclusion, this study provides empirical evidence that machine learning can be more efficient than the HVS in certain tasks but also highlights the strengths and limitations of both approaches. The authors suggest that combining CNNs with other classification techniques, such as bagof-words, may be a promising approach for improving the accuracy of dietary assessment through image food recognition. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.