Enhanced SOLOv2: An Effective Instance Segmentation Algorithm for Densely Overlapping Silkworms.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Sericulture Technology

Background:

• Accurate silkworm instance segmentation is vital for intelligent sericulture, enabling behavior analysis and health monitoring.
• High-density farming environments present challenges due to silkworm occlusion, hindering traditional segmentation methods.

Purpose of the Study:

• To develop an enhanced instance segmentation algorithm for accurately identifying individual silkworms in dense, occluded scenarios.
• To improve the reliability of biological parameter estimation in intelligent sericulture through precise segmentation.

Main Methods:

• Incorporation of Linear Deformable Convolution (LDC) for enhanced geometric feature modeling of curved silkworms.
• Integration of Haar Wavelet Downsampling (HWD) and an Edge-Augmented Multi-attention Fusion Network (EAMF-Net) to preserve details and improve boundary discrimination.
• Refinement of segmentation masks using Dynamic Upsampling (Dysample), Adaptive Spatial Feature Fusion (ASFF), and Simple Attention Module (SimAM).

Main Results:

• The enhanced SOLOv2 algorithm achieved an Average Precision (AP) of 85.1% on a self-built high-density silkworm dataset.
• Demonstrated significant improvements in segmentation accuracy for small targets (APs: +10.2%), medium targets (APm: +4.0%), and large targets (APl: +2.0%) compared to the baseline model.
• Effectively addressed challenges posed by high-density and severe mutual occlusion in silkworm farming environments.

Conclusions:

• The proposed enhanced SOLOv2 algorithm significantly advances precision in dense silkworm instance segmentation.
• This method provides a robust solution for individual silkworm analysis and health monitoring in intelligent sericulture.
• The improvements contribute to more reliable biological parameter estimation in real-world farming conditions.