Occlusion Robust Wheat Ear Counting Algorithm Based on Deep Learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate wheat ear counting is crucial for modern agriculture and crop yield evaluation.
• Dense wheat ear distribution causes significant occlusion and overlap issues in images.
• Traditional methods and existing deep learning models struggle to efficiently address occlusion in wheat ear counting.

Purpose of the Study:

• To develop an improved object detection model for accurate wheat ear counting, specifically addressing occlusion challenges.
• To enhance high-level semantic feature extraction for better wheat ear identification.
• To improve the efficiency and accuracy of automated wheat ear counting in agricultural settings.

Main Methods:

• Utilized transfer learning for pre-training the EfficientDet-D0 model's backbone to extract high-level semantic features.
• Proposed a novel image augmentation technique, Random-Cutout, to simulate occlusion by erasing image regions.
• Integrated the convolutional block attention module (CBAM) to refine features and focus on wheat ears, suppressing background noise.

Main Results:

• The improved EfficientDet-D0 model achieved a wheat ear counting accuracy of 94%, a 2% increase over the original model.
• The enhanced model demonstrated the lowest false detection rate at 5.8% compared to other methods.
• Experimental results confirmed the effectiveness of the proposed methods in handling occlusion and improving counting accuracy.

Conclusions:

• The improved EfficientDet-D0 model effectively addresses occlusion problems in wheat ear counting.
• The combination of transfer learning, Random-Cutout augmentation, and CBAM significantly boosts counting accuracy and reduces errors.
• This approach offers a promising solution for intelligent agriculture, enabling more precise crop yield estimation.