- You are here:
-
-
All publications
- Optimizing Data Collection Requirements of Image-Based Machine Learning Models for Improved Wild Blueberry (Vaccinium angustifolium Ait.) Ripeness Detection
All publications from your selection
Optimizing Data Collection Requirements of Image-Based Machine Learning Models for Improved Wild Blueberry (Vaccinium angustifolium Ait.) Ripeness Detection
Authors: Connor C. Mullins, Travis J. Esau, Qamar U. Zaman
Identifier: CSBE24276
Download file:
Published in: CSBE-SCGAB Technical Conferences » AGM Winnipeg 2024
Download RAW file:
DOI:
Description: This research outlined a comprehensive workflow to assess the viability of AI-generated imagery in training machine learning models for improving the detection of ripe wild blueberries (Vaccinium angustifolium Ait.). A dataset comprising of 200 high-resolution (26 MP) ground truth images of ripe wild blueberries were collected and augmented with AI-generated variations using DALLE 2 to increase overall dataset size. Models were then trained on three datasets: ground truth, generated, and a combination (40% of the dataset contribution being generated images). Evaluation metrics included precision, recall, mAP50, and mAP50-95, each analyzed using ANOVA multiple mean comparisons and Tukey?s HSD test through a completely randomized design. The results revealed that the ground truth models and the combination models had no significant difference across most performance metrics (p < 0.001) (mAP50, precision, and recall). The ground truth model achieved a mAP50 of 0.806, precision of 0.819, and recall of 0.723. The combination model achieved the highest mean performance across all metrics (mAP50: 0.834, precision: 0.854, recall: 0.755), with significantly higher performance on the mAP50-95 metric (0.478). This demonstrated the potential of AI-generated images to enhance training datasets. However, models trained solely on generated images showed significantly lower performance (mAP50: 0.642, mAP50-95: 0.308, precision: 0.743, recall: 0.566) when validated on ground truth images, indicating AI-generated images can augment datasets and improve generalization, but cannot fully replace ground truth data and maintain model performance. These findings highlighted the importance of a balanced approach to optimizing data collection protocols for wild blueberry ripeness detection.
Keywords: Image generation, DALLE 2, dataset augmentation, computer vision
Résumé:
Mots-clés:
Citation:
Volume:
Issue:
Pages -
Contributor:
Date: 2024-07-07
Technical field:
Conference name: CSBE/SCGAB 2023 Annual Conference, Winnipeg, Manitoba, 7-10 July 2024.
Session name: Refreshment Break & Poster Session
Other information:
Type: Poster
Format:
Publication type: Text.Abstract
Source:
Relation:
Coverage: North_America
Language 1: en
Language 2:
Rights: Canadian Society for Bioengineering
Notes:Department of Engineering, Faculty of Agriculture, Dalhousie University
Other files:
DOI:
Description: This research outlined a comprehensive workflow to assess the viability of AI-generated imagery in training machine learning models for improving the detection of ripe wild blueberries (Vaccinium angustifolium Ait.). A dataset comprising of 200 high-resolution (26 MP) ground truth images of ripe wild blueberries were collected and augmented with AI-generated variations using DALLE 2 to increase overall dataset size. Models were then trained on three datasets: ground truth, generated, and a combination (40% of the dataset contribution being generated images). Evaluation metrics included precision, recall, mAP50, and mAP50-95, each analyzed using ANOVA multiple mean comparisons and Tukey?s HSD test through a completely randomized design. The results revealed that the ground truth models and the combination models had no significant difference across most performance metrics (p < 0.001) (mAP50, precision, and recall). The ground truth model achieved a mAP50 of 0.806, precision of 0.819, and recall of 0.723. The combination model achieved the highest mean performance across all metrics (mAP50: 0.834, precision: 0.854, recall: 0.755), with significantly higher performance on the mAP50-95 metric (0.478). This demonstrated the potential of AI-generated images to enhance training datasets. However, models trained solely on generated images showed significantly lower performance (mAP50: 0.642, mAP50-95: 0.308, precision: 0.743, recall: 0.566) when validated on ground truth images, indicating AI-generated images can augment datasets and improve generalization, but cannot fully replace ground truth data and maintain model performance. These findings highlighted the importance of a balanced approach to optimizing data collection protocols for wild blueberry ripeness detection.
Keywords: Image generation, DALLE 2, dataset augmentation, computer vision
Résumé:
Mots-clés:
Citation:
Volume:
Issue:
Pages -
Contributor:
Date: 2024-07-07
Technical field:
Conference name: CSBE/SCGAB 2023 Annual Conference, Winnipeg, Manitoba, 7-10 July 2024.
Session name: Refreshment Break & Poster Session
Other information:
Type: Poster
Format:
Publication type: Text.Abstract
Source:
Relation:
Coverage: North_America
Language 1: en
Language 2:
Rights: Canadian Society for Bioengineering
Notes:Department of Engineering, Faculty of Agriculture, Dalhousie University
Other files:
©
2024
CSBE-SCGAB