A conditional segmentation-guided network for pomegranate image completion under occlusion.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Occlusion from leaves and branches in agricultural images hinders accurate pomegranate yield estimation and automated harvesting.
• Existing image completion methods struggle with structural fidelity in occluded agricultural imagery.

Purpose of the Study:

• To develop a novel framework for high-fidelity image completion and segmentation of occluded pomegranate fruits.
• To address the limitations of traditional methods in recovering structural integrity in agricultural images.

Main Methods:

• Proposed the Conditional Segmentation-guided Diffusion Network (CSD-Net), a lightweight, unified conditional diffusion model.
• Utilized a shared encoder, segmentation branch, and RGB diffusion branch.
• Leveraged segmentation masks as structural priors to guide the diffusion generation process for accurate reconstruction.

Main Results:

• CSD-Net achieved superior performance over conventional methods, with PSNR of 30.37 dB and SSIM of 0.9490.
• The model demonstrates high-fidelity reconstruction of fruit structures with spatial and textural consistency.
• Achieved a balance between high completion quality and inference efficiency with a model size of 117 MB.

Conclusions:

• CSD-Net offers a novel and effective solution for mitigating occlusion issues in agricultural visual perception.
• The proposed conditional guidance mechanism significantly improves structural integrity recovery in occluded pomegranate images.
• This work advances automated harvesting and yield estimation through enhanced visual perception in challenging agricultural conditions.