An Improved Instance Segmentation Method for Complex Elements of Farm UAV Aerial Survey Images

Area of Science:

• Agricultural Engineering
• Computer Vision
• Remote Sensing

Background:

• Traditional farm layer construction is inefficient and costly.
• Accurate segmentation of farm elements is crucial for unmanned operations like path planning and early warnings.

Purpose of the Study:

• To develop a high-precision instance segmentation algorithm for farm aerial imagery.
• To improve the efficiency and reduce the cost of creating farm survey layers.

Main Methods:

• Proposed a SparseInst-based algorithm incorporating a multi-scale attention module (MSA) with atrous convolution.
• Integrated a bottom-up aggregation path into the feature pyramid fusion network.
• Incorporated coordinate attention blocks (CAs) into the model's neck for enhanced semantic understanding.

Main Results:

• The improved SparseInst model achieved superior segmentation accuracy for complex farm targets compared to Mask R-CNN and Cascade-Mask.
• Demonstrated accuracy improvements of 10.8% and 12.8% over SOLOv2 and Condinst, respectively.
• The model requires the smallest number of parameters among the compared methods and supports real-time segmentation.

Conclusions:

• The proposed algorithm effectively segments diverse and complex targets in farm aerial imagery.
• This method offers a cost-effective and accurate solution for generating farm survey layers.
• The model is suitable for real-time applications in challenging farm environments.