DAM: Hierarchical Adaptive Feature Selection Using Convolution Encoder Decoder Network for Strawberry Segmentation.

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Autonomous harvesting of high-value crops like strawberries requires precise fruit identification.
• Real-time segmentation of strawberries is challenging due to occlusion from leaves, stems, and trusses in unbridled farming environments.

Purpose of the Study:

• To develop a dynamic feature selection mechanism for convolutional neural networks (CNNs) to improve strawberry segmentation.
• To enhance the accuracy and efficiency of strawberry detection for autonomous harvesting robots.

Main Methods:

• A novel dense attention module (DAM) was designed as a building block for CNNs, controlling information flow between encoder and decoder.
• DAM facilitates hierarchical adaptive feature fusion by analyzing inter-channel and intra-channel relationships.
• A custom dataset of strawberries across four maturity levels and background was created and utilized.

Main Results:

• The proposed DAM achieved a 4.1% increase in mean intersection over union (mIoU) compared to state-of-the-art semantic segmentation models.
• A 2.32% mIoU improvement was observed over existing attention modules.
• The method maintained a processing speed of 53 frames per second, ensuring real-time applicability.

Conclusions:

• The dense attention module (DAM) effectively improves semantic segmentation accuracy for strawberries in challenging agricultural settings.
• DAM can be easily integrated into existing CNN architectures, offering a versatile solution for agricultural robotics.
• The developed method enhances the potential for accurate autonomous strawberry harvesting by improving fruit ripeness detection.