A wheat spike detection method based on Transformer.

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Wheat spike detection is crucial for crop yield estimation and management.
• Convolutional Neural Networks (CNNs) face limitations in capturing global dependencies due to locality and scale-invariance.
• Transformers offer a promising alternative for extracting long-range dependencies in image data.

Purpose of the Study:

• To develop an advanced deep learning model for accurate wheat spike detection.
• To address the limitations of traditional CNNs in capturing global context.
• To improve wheat spike detection accuracy under complex field conditions.

Main Methods:

• Proposed a Transformer-based network, the Multi-Window Swin Transformer (MW-Swin Transformer), incorporating feature pyramid network capabilities for multi-scale feature extraction.
• Developed a novel Wheat Intersection over Union (IoU) loss function combining Euclidean distance, area overlap, and aspect ratio for improved bounding box regression.
• Integrated the MW-Swin Transformer and the new IoU loss into a fully convolutional one-stage object detection framework, named WheatFormer.
• Created and utilized the Wheat Spike Detection 2022 (WSD-2022) dataset for model evaluation.

Main Results:

• The WheatFormer model achieved a mean average precision (mAP) of 0.459 and an AP50 of 0.918.
• The proposed Transformer-based approach demonstrated superior performance compared to state-of-the-art algorithms.
• The novel IoU loss function contributed to enhanced detection accuracy.

Conclusions:

• The Transformer-based WheatFormer model is highly effective for wheat spike detection, even in challenging field environments.
• The integration of multi-scale feature extraction and an improved IoU loss function significantly boosts detection performance.
• This research provides a robust solution for precision agriculture applications in wheat production.