A lightweight wheat ear counting model in UAV images based on improved YOLOv8

Wheat (Triticum aestivum L.) is one of the significant food crops in the world, and the number of wheat ears serves as a critical indicator of wheat yield. Accurate quantification of wheat ear counts is crucial for effective scientific management of wheat fields. To address the challenges of missed detections, false detections, and diminished detection accuracy arising from the dense distribution, small size, and high overlap of wheat ears in Unmanned Aerial Vehicle (UAV) imagery, we propose a lightweight model, PSDS-YOLOv8 (P2-SPD-DySample-SCAM-YOLOv8), on the basis of the improved YOLOv8 framework, for the accurate detection of wheat ears in UAV images. First, the high resolution micro-scale detection layer (P2) is introduced to enhance the model's ability to recognize and localize small targets, while the large-scale detection layer (P5) is eliminated to minimize computational redundancy. Then, the Spatial Pyramid Dilated Convolution (SPD-Conv) module is employed to improve the ability of the network to learn features, thereby enhancing the representation of weak features of small targets and preventing information loss caused by low image resolution or small target sizes. Additionally, a lightweight dynamic upsampler, Dynamic Sample (DySample), is introduced to decrease computational complexity of the upsampling process by dynamically adjusting interpolation positions. Finally, the lightweight module Spatial Context-Aware Module (SCAM) is utilized to accurately map the connection between small targets and global features, enhancing the discrimination of small targets from the background. Experimental results demonstrate that the improved PSDS-YOLOv8 model achieves Mean Average Precision(mAP) 50 and mAP50:95 scores of 96.5% and 55.2%, which increases by 2.8% and 4.4%, while the number of parameters is reduced by 40.6% in comparison with the baseline YOLOv8 model. Compared to YOLOv5, YOLOv7, YOLOv9, YOLOv10, YOLOv11, Faster RCNN, SSD, and RetinaNet, the improved model demonstrates superior accuracy and fewer parameters, exhibiting the best overall performance. The methodology proposed in this study enhances model accuracy while concurrently reducing resource consumption and effectively addressing the issues of missed and false detections of wheat ears, thereby providing technical support and theoretical guidance for intelligent counting of wheat ears in UAV imagery.