A Multiscale Point-Supervised Network for Counting Maize Tassels in the Wild.

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Accurate maize tassel counting is crucial for crop monitoring and yield estimation.
• Existing deep learning object detection methods struggle with varying tassel scales and complex backgrounds, leading to inefficient detection.
• Current methods often require bounding box annotations, which are labor-intensive.

Purpose of the Study:

• To propose a novel deep learning model, MLAENet, for accurate maize tassel counting using only point-level annotations.
• To address the challenges of scale variation and background complexity in maize tassel detection.
• To improve the efficiency and accuracy of automated crop monitoring systems.

Main Methods:

• Developed a multiscale lite attention enhancement network (MLAENet) utilizing point-level annotations.
• Introduced a multicolumn lite feature extraction module with dilated convolutions for scale-dependent density map generation.
• Implemented a multifeature enhancement module with an attention strategy to handle complex backgrounds.
• Designed an UP-Block module to suppress gridding effects during density map up-sampling.

Main Results:

• MLAENet demonstrated significant advantages in counting accuracy compared to state-of-the-art methods on two public datasets.
• The proposed network achieved improved inference speed.
• The model effectively handles scale variations and complex backgrounds in maize tassel images.

Conclusions:

• MLAENet offers a robust and efficient solution for maize tassel counting using point annotations.
• The method shows potential for enhancing precision agriculture and crop yield prediction.
• The model's public availability facilitates further research and application in crop monitoring.