Wheat Seed Detection and Counting Method Based on Improved YOLOv8 Model.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate wheat seed counting is crucial for agricultural applications like calculating thousand-grain weight and crop breeding.
• Existing methods struggle with seed accumulation, adhesion, and occlusion, leading to low accuracy and speed.
• Developing efficient embedded systems for seed counters requires optimized detection models.

Purpose of the Study:

• To propose a lightweight, real-time wheat seed detection model (YOLOv8-HD) for improved counting accuracy and speed.
• To address challenges posed by clustered, adhered, and occluded wheat seeds in detection.
• To provide technical support for the development of embedded seed counting platforms.

Main Methods:

• Introduced shared convolutional layers in the YOLOv8 detection head for a lightweight design and improved speed.
• Integrated Vision Transformer with Deformable Attention into the backbone's C2f module to enhance feature extraction.
• Developed the YOLOv8-HD model based on the YOLOv8 architecture.

Main Results:

• YOLOv8-HD achieved 77.6% mAP in stacked scenes with impurities, a 9.1% improvement over YOLOv8.
• Overall mAP reached 99.3%, a 16.8% increase compared to YOLOv8.
• Model size reduced to 6.35 MB (4/5 of YOLOv8), GFLOPs decreased by 16%, and inference time was 2.86 ms (on GPU).

Conclusions:

• YOLOv8-HD demonstrates superior performance in wheat seed detection compared to mainstream networks regarding accuracy, speed, and model size.
• The model efficiently detects wheat seeds in diverse scenarios, including challenging conditions like severe adhesion.
• YOLOv8-HD provides effective technical support for developing advanced seed counting instruments.