YOLOv11-SMS: An Improved Algorithm for Impurity Detection in Seed Cotton.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Current cottonseed impurity detection methods suffer from high miss-detection rates and suboptimal performance.
• There is a need for enhanced precision and efficiency in automated agricultural quality control systems.

Purpose of the Study:

• To introduce an improved YOLOv11 algorithm (YOLOv11-SMS) for precise cottonseed impurity detection.
• To enhance the model's ability to learn local and global features, improve feature extraction, and optimize small target detection.

Main Methods:

• Integration of a local self-attention mechanism (LRSA) within the C2PSA-SL module for augmented local information learning.
• Incorporation of a multi-branch reparameterized convolution (MBRConv) module for enhanced feature extraction while maintaining lightweight properties.
• Introduction of a spatial adaptive modulation (SAFM) module to optimize the detection of small targets.

Main Results:

• YOLOv11-SMS demonstrated improved performance over the baseline model, with mAP@50-95 increasing by 3.07 percentage points (from 79.42% to 82.49%).
• The average mIOU saw a 3.2 percentage point improvement, rising from 90.98% to 94.18%.
• The model achieved a real-time inference speed of 178.63 frames per second (FPS).

Conclusions:

• YOLOv11-SMS effectively balances detection accuracy and speed, offering an efficient and precise solution for cottonseed impurity detection.
• The proposed enhancements significantly improve the performance of the YOLOv11 algorithm for agricultural quality control applications.