A New Semantic Segmentation Framework Based on UNet.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Agricultural robots require advanced environmental perception for autonomous operation.
• Semantic segmentation is crucial for identifying objects and navigable areas in agricultural settings.
• Integrating multiple data sources and models can improve segmentation accuracy.

Purpose of the Study:

• To develop and evaluate an ensemble semantic segmentation framework for agricultural intelligence.
• To enhance the environmental awareness capabilities of agricultural robots.
• To assess the framework's performance against existing state-of-the-art methods.

Main Methods:

• An ensemble framework was developed using the bagging strategy and the UNet network.
• The framework integrates RGB and HSV color spaces for improved feature representation.
• Performance was evaluated on self-built (Maize) and public (Sugar Beets) datasets.

Main Results:

• The ensemble framework achieved the highest Intersection over Union (IoU) scores: 0.8276 on Maize and 0.6972 on Sugar Beets.
• Compared to single-color space UNet methods, the ensemble approach effectively synthesized advantages.
• UNet-based methods, including the ensemble, demonstrated faster processing speeds and smaller parameter spaces than DeepLab V3+ and SegNet.

Conclusions:

• The proposed ensemble semantic segmentation framework offers a robust solution for agricultural intelligence.
• The framework's efficiency and accuracy make it suitable for deployment on resource-constrained agricultural robots.
• This approach contributes to the advancement of autonomous agricultural systems.