EasyDAM_V3: Automatic Fruit Labeling Based on Optimal Source Domain Selection and Data Synthesis via a Knowledge...

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Deep learning for fruit detection requires extensive labeled data, which is costly and time-consuming to acquire.
• Existing methods for reducing labeling costs, like generative adversarial networks, have limitations in cross-species applicability and fully automating the process.

Purpose of the Study:

• To develop an improved automatic labeling method (EasyDAM_V3) for fruit detection, eliminating manual labeling and reducing costs.
• To establish an optimal source domain selection method using a multidimensional spatial feature model.
• To create a high-volume dataset construction method utilizing transparent background fruit image translation and knowledge graphs.

Main Methods:

• Proposed the EasyDAM_V3 model for automatic fruit label acquisition.
• Developed an optimal source domain selection strategy based on multidimensional spatial features.
• Implemented a dataset construction method involving transparent background fruit image translation and knowledge graph synthesis.

Main Results:

• The EasyDAM_V3 model successfully automated fruit labeling, eliminating the need for manual annotation.
• Achieved high average annotation precision: 90.94% for orange, 89.78% for apple, and 90.84% for tomato.
• Demonstrated the model's capability to identify the optimal source domain (pear) for target datasets.

Conclusions:

• The EasyDAM_V3 model effectively automates fruit labeling tasks, significantly reducing manual labor and associated costs.
• The proposed methods for optimal source domain selection and dataset construction enhance the efficiency and accuracy of automatic labeling.
• This approach makes deep learning-based fruit detection more feasible by overcoming data annotation bottlenecks.