Ikuko Shimizu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ikuko Shimizu has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Rebeka Sultana
Tokyo University of Agriculture and Technology, Tokyo, Japan
8 shared publications - Hironori Nakajo
Tokyo University of Agriculture and Technology
3 shared publications - Akio Sashima
National Institute of Advanced Industrial Science and Technology, Kashiwa, Japan
2 shared publications Hibiki Horibe
2 shared publications- Ikuma Esaki
Tokyo University of Agriculture and Technology, 2-24-16 Nakachō, Tokyo 184-0012, Japan
2 shared publications - Rinka Kiriyama
Department of Electrical Engineering and Computer Science, Tokyo University of Agriculture and Technology, 2-24-16 Naka-cho, Koganei, Tokyo 162-0061, Japan
2 shared publications
Latest Publications
- Open Access
- 25 May 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (6) , 223
DiSMix: Dimensional Swap Mix for Feature-Level Data Augmentation in Vision Transformers
Mixup is a data augmentation technique that improves prediction accuracy in classification tasks by combining representations of training samples, which makes it particularly effective in settings with limited data and during fine-tuning for downstream tasks. However, representations generated by mixup may appear unnatural, which can negatively affect fine-tuning performance. To address this limitation, we propose a vision transformer (ViT)-aware variant of mixup strategies, Dimensional Swap Mix (DiSMix). DiSMix divides a representation vector into two segments corresponding to subspaces of the original feature space and generates new representations by swapping one segment with that from another sample and concatenating the segments. This allows part of the original representation to remain unchanged, enabling the model to learn from partially preserved features. We evaluate DiSMix by applying several mixup-based methods to fine-tune ViTs on the VTAB-1k benchmark. The findings show that DiSMix improves accuracy on the VTAB-1k Natural split, reaching 80.0%, compared with conventional mixup methods. This suggests that DiSMix is an effective alternative for representation-level data augmentation in fine-tuning scenarios.
- 25 May 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (6) , 223
Open AccessDiSMix: Dimensional Swap Mix for Feature-Level Data Augmentation in Vision Transformers
Mixup is a data augmentation technique that improves prediction accuracy in classification tasks by combining representations of training samples, which makes it particularly effective in settings with limited data and during fine-tuning for downstream tasks. However, representations generated by mixup may appear unnatural, which can negatively affect fine-tuning performance. To address this limitation, we propose a vision transformer (ViT)-aware variant of mixup strategies, Dimensional Swap Mix (DiSMix). DiSMix divides a representation vector into two segments corresponding to subspaces of the original feature space and generates new representations by swapping one segment with that from another sample and concatenating the segments. This allows part of the original representation to remain unchanged, enabling the model to learn from partially preserved features. We evaluate DiSMix by applying several mixup-based methods to fine-tune ViTs on the VTAB-1k benchmark. The findings show that DiSMix improves accuracy on the VTAB-1k Natural split, reaching 80.0%, compared with conventional mixup methods. This suggests that DiSMix is an effective alternative for representation-level data augmentation in fine-tuning scenarios.
- Open Access
- 15 April 2026
- Published by Elsevier in Informatics in Medicine Unlocked
- Vol. 63
Cancer cytoplasm segmentation in hyperspectral cell image with data augmentation
- 15 April 2026
- Published by Elsevier in Informatics in Medicine Unlocked
- Vol. 63
Open AccessCancer cytoplasm segmentation in hyperspectral cell image with data augmentation
- Open Access
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42569-42591
Stacking Binary Neural Networks for Efficient Edge-Based Waste Classification
Waste classification poses significant challenges owing to rapid societal development and increasing waste generation. Although deep neural networks have demonstrated remarkable results in automating this complex task, scaling deployment to household-level applications remains constrained by hardware limitations. This study enhanced Binary Neural Network architectures by integrating a stacking ensemble approach and evaluated its performance on Advanced RISC Machine (ARM)-based edge devices using the Larq Compute Engine (LCE). The stacking strategy combines three diverse binary architectures, ResNet18, DenseNet121, and ShuffleNetV2, trained with adaptive alpha scaling and Parametric ReLU activations and integrated through a lightweight binary meta-learner. We expanded the publicly available Kaggle dataset from 15,150 to 21,416 images (a 41% increase) across the 12 waste categories. Through a comprehensive evaluation with 5-fold cross-validation, our ensemble achieved a 91.42% and 4.6 percentage point improvement over our previous single-model DRA-BNN. More specifically, real-world deployment evaluated on a Raspberry Pi 3B+ demonstrates a throughput of 6.24 images/s in offline batch testing and 3.24 images/s during continuous 1-hour operation, while maintaining stable performance at 82% average CPU load and 72.2 °C average CPU temperature. Compared to full-precision ResNet18, our method achieves 2.65×faster throughput in offline testing (6.24 vs. 2.35 img/s) and 3.09×improvement in real-time inference (3.24 vs. 1.05 img/s), while maintaining an accuracy gap of only 2.8%. The Larq-optimized deployment achieved a 32× reduction in memory usage and utilized ARM instructions for efficient binary operations. These results demonstrate effective accuracy–efficiency trade-offs for practical edge-based waste sorting.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42569-42591
Open AccessStacking Binary Neural Networks for Efficient Edge-Based Waste Classification
Waste classification poses significant challenges owing to rapid societal development and increasing waste generation. Although deep neural networks have demonstrated remarkable results in automating this complex task, scaling deployment to household-level applications remains constrained by hardware limitations. This study enhanced Binary Neural Network architectures by integrating a stacking ensemble approach and evaluated its performance on Advanced RISC Machine (ARM)-based edge devices using the Larq Compute Engine (LCE). The stacking strategy combines three diverse binary architectures, ResNet18, DenseNet121, and ShuffleNetV2, trained with adaptive alpha scaling and Parametric ReLU activations and integrated through a lightweight binary meta-learner. We expanded the publicly available Kaggle dataset from 15,150 to 21,416 images (a 41% increase) across the 12 waste categories. Through a comprehensive evaluation with 5-fold cross-validation, our ensemble achieved a 91.42% and 4.6 percentage point improvement over our previous single-model DRA-BNN. More specifically, real-world deployment evaluated on a Raspberry Pi 3B+ demonstrates a throughput of 6.24 images/s in offline batch testing and 3.24 images/s during continuous 1-hour operation, while maintaining stable performance at 82% average CPU load and 72.2 °C average CPU temperature. Compared to full-precision ResNet18, our method achieves 2.65×faster throughput in offline testing (6.24 vs. 2.35 img/s) and 3.09×improvement in real-time inference (3.24 vs. 1.05 img/s), while maintaining an accuracy gap of only 2.8%. The Larq-optimized deployment achieved a 32× reduction in memory usage and utilized ARM instructions for efficient binary operations. These results demonstrate effective accuracy–efficiency trade-offs for practical edge-based waste sorting.
- Open Access
- 21 October 2025
- Published by MDPI AG in Horticulturae
- Vol. 11 (10) , 1272
Maturity Classification of Blueberry Fruit Using YOLO and Vision Transformer for Agricultural Assistance
This paper proposes a method for classifying the maturity levels of blueberry fruit from camera images as part of a cultivation support system. Following the five-stage maturity classification, the proposed approach first detects individual blueberry regions in an image and subsequently classifies each region into one of the five levels. The method leverages a Transformer-based model to extract features from local fruit regions that include contextual background, enabling the learning of spatial relationships both within and beyond the fruit boundaries. A dedicated dataset was constructed by capturing images of blueberry fruits alongside a color chart representing maturity levels. Experimental evaluations involving multiple deep learning models under three training–testing configurations demonstrate the effectiveness of the proposed method, achieving an average classification accuracy of 93.7%.
- 21 October 2025
- Published by MDPI AG in Horticulturae
- Vol. 11 (10) , 1272
Open AccessMaturity Classification of Blueberry Fruit Using YOLO and Vision Transformer for Agricultural Assistance
This paper proposes a method for classifying the maturity levels of blueberry fruit from camera images as part of a cultivation support system. Following the five-stage maturity classification, the proposed approach first detects individual blueberry regions in an image and subsequently classifies each region into one of the five levels. The method leverages a Transformer-based model to extract features from local fruit regions that include contextual background, enabling the learning of spatial relationships both within and beyond the fruit boundaries. A dedicated dataset was constructed by capturing images of blueberry fruits alongside a color chart representing maturity levels. Experimental evaluations involving multiple deep learning models under three training–testing configurations demonstrate the effectiveness of the proposed method, achieving an average classification accuracy of 93.7%.
- Open Access
Cancer cytoplasm segmentation in hyperspectral cell image with data augmentation
Hematoxylin and Eosin (H&E)-stained images are commonly used to detect nuclear or cancerous regions in cells from images captured by a microscope. Identifying cancer cytoplasm is crucial for determining the type of cancer; hence, obtaining accurate cancer cytoplasm regions in cell images is important. While CMOS images often lack detailed information necessary for diagnosis, hyperspectral images provide more comprehensive cell information. Using a deep learning model, we propose a method for detecting cancer cell cytoplasm in hyperspectral images. Deep learning models require large datasets for learning; however, capturing a large number of hyperspectral images is difficult. Additionally, hyperspectral images frequently contain instrumental noise, depending on the characteristics of the imaging devices. We propose a data augmentation method to account for instrumental noise. CMOS images were used for data augmentation owing to their visual clarity, which facilitates manual annotation compared to original hyperspectral images. Experimental results demonstrate the effectiveness of the proposed data augmentation method both quantitatively and qualitatively.
Open AccessCancer cytoplasm segmentation in hyperspectral cell image with data augmentation
Hematoxylin and Eosin (H&E)-stained images are commonly used to detect nuclear or cancerous regions in cells from images captured by a microscope. Identifying cancer cytoplasm is crucial for determining the type of cancer; hence, obtaining accurate cancer cytoplasm regions in cell images is important. While CMOS images often lack detailed information necessary for diagnosis, hyperspectral images provide more comprehensive cell information. Using a deep learning model, we propose a method for detecting cancer cell cytoplasm in hyperspectral images. Deep learning models require large datasets for learning; however, capturing a large number of hyperspectral images is difficult. Additionally, hyperspectral images frequently contain instrumental noise, depending on the characteristics of the imaging devices. We propose a data augmentation method to account for instrumental noise. CMOS images were used for data augmentation owing to their visual clarity, which facilitates manual annotation compared to original hyperspectral images. Experimental results demonstrate the effectiveness of the proposed data augmentation method both quantitatively and qualitatively.
- Open Access
- 2 July 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (13) , 7440
Ensemble Learning-Based Weed Detection from a Duck’s Perspective Using an Aquatic Drone in Rice Paddies
Semantic segmentation using neural networks (NNs) has significant potential for weed detection in agricultural fields. However, conventional datasets captured from aerial perspectives often fail to detect weeds that are either hidden beneath crops or submerged in water. This study proposes a method for accurately detecting weed pixels through ensemble learning-based semantic segmentation, using forward-facing images captured by a camera mounted on an aquatic drone navigating between rice plants. We also present a paddy field weed image dataset constructed to train the NN models. Multiple semantic segmentation models were trained, compared, and evaluated, achieving a weed intersection over union (IoU) of 0.441, mean IoU (mIoU) of 0.706, and pixel accuracy of 0.971.
- 2 July 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (13) , 7440
Open AccessEnsemble Learning-Based Weed Detection from a Duck’s Perspective Using an Aquatic Drone in Rice Paddies
Semantic segmentation using neural networks (NNs) has significant potential for weed detection in agricultural fields. However, conventional datasets captured from aerial perspectives often fail to detect weeds that are either hidden beneath crops or submerged in water. This study proposes a method for accurately detecting weed pixels through ensemble learning-based semantic segmentation, using forward-facing images captured by a camera mounted on an aquatic drone navigating between rice plants. We also present a paddy field weed image dataset constructed to train the NN models. Multiple semantic segmentation models were trained, compared, and evaluated, achieving a weed intersection over union (IoU) of 0.441, mean IoU (mIoU) of 0.706, and pixel accuracy of 0.971.
- Open Access
- 4 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 96-104
Two-shot Spatially Varying Bidirectional Reflectance Distribution Function Estimation for Lustrous Material by Deep Learning
- 4 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 96-104
Open AccessTwo-shot Spatially Varying Bidirectional Reflectance Distribution Function Estimation for Lustrous Material by Deep Learning
- Open Access
- 4 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 75-79
Maturity classification of blueberry fruit from camera image for cultivation support system
- 4 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 75-79
Open AccessMaturity classification of blueberry fruit from camera image for cultivation support system
- Open Access
- 14 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 189-193
Enhancing Daytime Train Image Detection with YOLOv8 through Data Augmentation Techniques
- 14 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 189-193
Open AccessEnhancing Daytime Train Image Detection with YOLOv8 through Data Augmentation Techniques
- Open Access
- 14 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 237-243
Empirical Analysis of Image-Based 3D Reconstruction Technique for Museum Artifacts Incorporating Artificial Features
- 14 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 237-243
Open AccessEmpirical Analysis of Image-Based 3D Reconstruction Technique for Museum Artifacts Incorporating Artificial Features