BiSeNeXt: a yam leaf and disease segmentation method based on an improved BiSeNetV2 in complex scenes

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Yam quality and yield are significantly impacted by leaf diseases, necessitating accurate monitoring.
• Current research on yam leaf disease segmentation is limited, facing challenges like overlapping leaves, uneven lighting, and irregular disease spots.
• Accurate segmentation is crucial for disease identification and management in yam cultivation.

Purpose of the Study:

• To develop the first dataset for yam leaf disease segmentation.
• To propose an enhanced segmentation method, BiSeNeXt, for improved accuracy in complex environments.
• To provide a robust foundation for analyzing yam leaf health and diseases.

Main Methods:

• Introduced the first yam leaf disease segmentation dataset.
• Developed BiSeNeXt, an enhanced segmentation method based on BiSeNetV2.
• Incorporated Dynamic Feature Extraction Block (DFEB) with Dynamic Receptive-Field Convolution (DRFConv) and Pixel Shuffle (PixelShuffle) for precise edge detection.
• Utilized Efficient Asymmetric Multi-Scale Attention (EAMA) to address lesion adhesion.
• Employed PointRefine decoder for adaptive refinement of segmentation predictions.

Main Results:

• Achieved 97.04% Intersection over Union (IoU) for leaf segmentation and 84.75% IoU for disease segmentation.
• Improved IoU by 2.22% for leaf segmentation and 5.58% for disease segmentation compared to DeepLabV3+.
• Demonstrated significantly lower computational cost, requiring only 11.81% of FLOPs and 7.81% of parameters compared to DeepLabV3+.

Conclusions:

• BiSeNeXt accurately and efficiently segments yam leaf spots in complex scenes.
• The proposed method offers a significant advancement for yam disease analysis and management.
• This work establishes a strong foundation for further research in agricultural image analysis.