High-Throughput Spike Detection in Greenhouse Cultivated Grain Crops with Attention Mechanisms-Based Deep Learning...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop yield assessment relies on detecting grain spikes, which are often obscured by leaves.
• Existing deep neural networks (DNNs) struggle with precise spike detection due to their small size and similar appearance to foliage.

Purpose of the Study:

• To improve the accuracy and efficiency of grain spike detection in images.
• To introduce architectural modifications to the Faster R-CNN (FRCNN) model, incorporating a global attention module.

Main Methods:

• Modified the Faster R-CNN architecture by reducing feature extraction layers and adding a global attention module, creating FRCNN-A.
• Evaluated FRCNN-A against conventional FRCNN and Swin Transformer on diverse European wheat cultivars, including challenging phenotypes.
• Tested models on both baseline and FastGAN-augmented datasets to assess performance under varying data conditions.

Main Results:

• FRCNN-A demonstrated improved detection accuracy for inner spikes, achieving a mean average precision (mAP) of 81.0% compared to FRCNN's 76.0%.
• FRCNN-A showed competitive performance against the Swin Transformer (83.0% mAP) on inner spikes.
• FRCNN-A proved to be a faster, lightweight network than FRCNN and Swin Transformer across datasets.
• On augmented data, FRCNN-A reached 85.0% mAP, outperforming FRCNN (84.24%) and approaching Swin Transformer (89.45%).

Conclusions:

• Architectural adaptations, particularly attention mechanisms, significantly enhance DNN performance in detecting subtle grain spike features.
• FRCNN-A offers a promising, efficient solution for image-based quantitative crop yield assessment.