A novel multiscale feature enhancement network using learnable density map for red clustered pepper yield estimation.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate red cluster pepper (RCP) yield estimation is crucial for optimizing agricultural management and resource allocation.
• Traditional methods face challenges with labor-intensive annotation and reduced accuracy in dense crop settings.

Purpose of the Study:

• To develop an automated and accurate yield estimation method for RCP.
• To overcome limitations of traditional object detection methods in dense environments.

Main Methods:

• Proposed a novel multiscale feature enhancement network (MFEN) integrated with a learnable density map (LDM).
• Enhanced the kernel-based density map (KDM) method using Swin Transformer (ST) to create LDM.
• MFEN incorporates dilation convolution, residual structures, and attention mechanisms for feature extraction.
• Jointly trained LDM and MFEN for simultaneous yield and density map estimation.

Main Results:

• The MFEN with LDM achieved superior accuracy in RCP yield estimation.
• Achieved a R-squared value of 0.9802 on the test dataset, outperforming DSNet by 0.98%.
• Demonstrated efficient deployment capabilities with 13.08M parameters, significantly fewer than CSRNet.

Conclusions:

• The proposed MFEN and LDM integration offers a robust algorithmic solution for intelligent RCP yield estimation.
• The method enhances accuracy and efficiency in automated agricultural yield prediction.
• Provides a foundation for advanced precision agriculture practices.