SRNet-YOLO: A model for detecting tiny and very tiny pests in cotton fields based on super-resolution reconstruction

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Effective pest management is crucial for cotton cultivation.
• Tiny pests (smaller than 32x32 pixels) and very tiny pests (smaller than 16x16 pixels) pose detection challenges for conventional object detection models.
• Accurate detection of these small pests is vital for informed agricultural decisions.

Purpose of the Study:

• To develop an advanced framework, SRNet-YOLO, for accurately detecting tiny and very tiny pests in wild cotton fields.
• To enhance the feature extraction and fusion capabilities for improved small object detection.
• To create a comprehensive dataset for validating the proposed method.

Main Methods:

• Proposed SRNet-YOLO framework integrating YOLOv8 feature extraction, a feature map super-resolution reconstruction module (FM-SR), and a BiFormer attention-based fusion mechanism (BiFormerAF).
• The FM-SR module reconstructs P5 layer feature maps to P3 layer size to recover detailed features.
• The BiFormerAF module fuses reconstructed features with P3 layer features to prevent information loss and improve detection.

Main Results:

• SRNet-YOLO achieved 78.2% mAP for tiny pests, outperforming YOLOv3, YOLOv5, YOLOv7, and YOLOv8.
• For very tiny pests, the framework reached 57% mAP, a 32.2% improvement over YOLOv8.
• The model demonstrated strong generalizability, achieving 92.8% mAP on the Yellow Sticky Traps dataset.

Conclusions:

• SRNet-YOLO effectively addresses the challenge of detecting tiny pests in cotton fields.
• The proposed framework exhibits excellent generalizability, applicable to tiny pest detection in other crops.
• This research contributes a robust solution for precision agriculture through enhanced computer vision techniques.