CNATNet: a convolution-attention hybrid network for safflower classification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Safflower (Carthamus tinctorius L.) is a vital crop for medicinal and economic uses.
• Manual filament grading is labor-intensive, slow, and not scalable for quality control.
• Accurate grading is crucial for agricultural and pharmaceutical applications of safflower.

Purpose of the Study:

• To develop an efficient and accurate automated system for safflower filament grading.
• To establish a coarse-to-fine grading framework for rapid assessment and fine-grained classification.
• To design a lightweight hybrid network suitable for real-time, resource-constrained environments.

Main Methods:

• A novel lightweight hybrid network, CNATNet, integrating convolutional operations and attention mechanisms was developed.
• The network features optimized components: C2S2 (lightweight convolutional variant) and AnC2f (n-order local attention).
• A depthwise separable convolution-based head (DWClassify) was used for accelerated inference.

Main Results:

• CNATNet achieved high accuracy: 98.6% at the cluster level and 95.6% at the filament level.
• The system demonstrated low latency (1.9 ms per image) and real-time performance (63 FPS on Jetson Orin Nano).
• CNATNet outperformed baseline models like YOLOv11m and RT-DETRv2s in accuracy and speed.

Conclusions:

• CNATNet offers a task-specific, lightweight solution for safflower filament grading, balancing accuracy and efficiency.
• The developed framework is feasible for practical, embedded agricultural quality classification in resource-limited settings.
• This automated approach holds significant potential for improving safflower quality control in industry.