Low-Cost Optimized U-Net Model with GMM Automatic Labeling Used in Forest Semantic Segmentation

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) are crucial for image analysis tasks like classification and segmentation.
• Training state-of-the-art CNNs requires large, well-labeled datasets, which are often costly and time-consuming to create manually.
• Manual data labeling is a significant bottleneck in developing efficient deep learning models for image analysis.

Purpose of the Study:

• To propose an automated method for generating training labels for Convolutional Neural Networks (CNNs).
• To optimize the U-Net model hyperparameters for a balance between performance and structural complexity.
• To develop a low-cost, efficient system for image analysis applications.

Main Methods:

• Implemented an unsupervised clustering technique using Gaussian Mixture Models (GMM) for automatic label generation.
• Optimized hyperparameters of the U-Net model to enhance efficiency and reduce complexity.
• Applied and tested the developed methods on a deforestation monitoring task using aerial imagery.

Main Results:

• The proposed GMM-based method significantly reduced the resources and computation time required for data labeling.
• Optimized U-Net model achieved a balance between high accuracy and reduced model complexity.
• The system demonstrated effectiveness in identifying forests in aerial imagery for deforestation monitoring.

Conclusions:

• Automated label generation using GMMs offers a viable alternative to manual labeling for CNN training.
• Hyperparameter optimization of U-Net models can lead to more efficient and cost-effective deep learning systems.
• The developed approach successfully addresses challenges in resource-intensive image analysis tasks like deforestation monitoring.