DCP-YOLOv7x: improved pest detection method for low-quality cotton image

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate detection of cotton pests is crucial for crop yield and quality.
• Low-light conditions pose significant challenges to current pest detection methods due to degraded image quality and difficult feature extraction.

Purpose of the Study:

• To develop an effective method for detecting cotton pests in low-light environments.
• To address the limitations of existing methods in terms of image quality, feature extraction, and detection precision.

Main Methods:

• The proposed DCP-YOLOv7x method utilizes FFDNet and EnlightenGAN for low-light image enhancement.
• Incorporates a Deformable Attention (DAttention) mechanism and a Dynamic Head (DyHead) structure for improved feature extraction.
• Modifies the loss function with Normalized Wasserstein Distance (NWD) to enhance small target detection.

Main Results:

• DCP-YOLOv7x achieved a detection Precision (P) of 95.9% and mAP@0.5 of 95.4% for cotton pests in low-light conditions.
• Demonstrated significant improvements of 14.4% in Precision and 15.6% in mAP@0.5 compared to the standard YOLOv7x.
• Validated effectiveness across different low-light datasets, including Exdark and Dk-CottonInsect.

Conclusions:

• DCP-YOLOv7x offers a robust solution for accurate and efficient cotton pest detection in challenging low-light environments.
• The method provides a strong theoretical foundation for enhancing cotton quality and yield.
• Potential for integration into agricultural edge computing devices for practical field applications.