TasselNetv2

TasselNetv2 counts wheat spikes in field images to provide quantitative measurements for crop-yield-related phenotyping.

Key Features:

• Context-Augmented Local Regression Networks: Incorporates visual context into the analysis of local image patches to improve spike counts when spikes are partially visible.
• Efficiency and Speed: Implemented in a fully convolutional form to reduce redundant computations during training and inference, achieving 13.82 frames per second on 912 × 1216 images.
• High Accuracy: Reports 91.01% counting accuracy on the Wheat Spikes Counting (WSC) dataset comprising 1,764 high-resolution images and 675,322 manually-annotated instances.
• Versatility Across Datasets: Advances performance on the Maize Tassels Counting and ShanghaiTech Crowd Counting datasets.
• Scalability and Adaptability: Achieves accurate results when trained from scratch with small networks while larger pre-trained networks can further enhance accuracy.

Scientific Applications:

• Crop yield estimation: Provides quantitative spike counts to support phenotyping, crop management decisions, and breeding program evaluations.
• Object counting in computer vision: Applies to other counting tasks such as maize tassel counting and crowd counting for high-throughput quantitative analyses.

Methodology:

Uses convolutional neural networks with context-augmented local regression networks implemented in a fully convolutional architecture to incorporate contextual information while reducing redundant computations.