Publicação
The role of background colour in pollen recognition task using CNN
| Resumo: | Pollen recognition is a crucial but challenging task in addressing a variety of questions like pollination or palaeobotany, but also for other fields of research, e.g., allergology, melissopalynology or forensics. State-of-the-art methods mainly use deep learning CNNs for pollen recognition, however, we observe that existing published approaches use original images without study the possible biased recognition due to pollen’s background colour. In this paper, we evaluate the DenseNet model trained with original images and with segmented images (remove background) and analyse network’s predictive performance under these conditions using a cross evaluation approach. An accuracy of 97.4% was achieved that represents one of the best successes rate when weighted with the number of taxa of any attempt at automated pollen analysis currently documented in the literature. From these results, we confirm the existence of background specific influence in the recognition task. |
|---|---|
| Autores principais: | Monteiro, Fernando C. |
| Outros Autores: | Pinto, Cristina M.; Rufino, José |
| Assunto: | Pollen recognition Deep learning Convolutional neural network |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | documento de conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Instituto Politécnico de Bragança |
| Idioma: | inglês |
| Origem: | Biblioteca Digital do IPB |
| Resumo: | Pollen recognition is a crucial but challenging task in addressing a variety of questions like pollination or palaeobotany, but also for other fields of research, e.g., allergology, melissopalynology or forensics. State-of-the-art methods mainly use deep learning CNNs for pollen recognition, however, we observe that existing published approaches use original images without study the possible biased recognition due to pollen’s background colour. In this paper, we evaluate the DenseNet model trained with original images and with segmented images (remove background) and analyse network’s predictive performance under these conditions using a cross evaluation approach. An accuracy of 97.4% was achieved that represents one of the best successes rate when weighted with the number of taxa of any attempt at automated pollen analysis currently documented in the literature. From these results, we confirm the existence of background specific influence in the recognition task. |
|---|