LKD-YOLO: Improved YOLOv8 for Rice panicle detection in drone images

Rice is one of the major staple crops in the world, and accurately and quickly identifying rice panicles is a key step in implementing precision agriculture in the rice sector. The rise of computer vision technology and drone technology has provided effective means for rice panicle detection. However, in drone images, rice panicles occupy a small portion of the image pixels and are densely distributed, often overlapping with each other, and the features differ depending on the height of the capture. These challenges make it difficult for existing object detection models to detect rice panicles effectively. This paper primarily uses YOLOv8 as the baseline model and proposes the LKD-YOLO detection method. LKD-YOLO introduces an additional small target detection layer, which enhances the model's ability to detect small targets. To address issues in the dataset such as multi-height distribution, occlusion, and density, we proposed the LKM module, a feature extraction module that primarily uses multi-scale large kernel convolutions. This module effectively extracts key features of rice panicles. Additionally, we improved the detection head by incorporating an attention mechanism-based target detection head, known as Dyhead. As a result, the final model's detection accuracy, measured by mAP50, improved by $3.6 \%$ compared to the baseline model (YOLOv8s).