Nowadays, with artificial intelligence quickly developing, computer vision is being further applied across various areas. In agriculture, computer vision has become a valuable tool for seed counting tasks, offering higher efficiency and saving resources compared to traditional manual counting methods. For Pinus tabuliformis seeds, their small size and dense distribution lead to low detection efficiency and difficulty in achieving accurate identification. To address this issue, an improved lightweight YOLOv8-based detection algorithm is proposed. The algorithm introduces an asymmetric decoupled detection head (LADH) to optimize the YOLOv8′s head, which expands the receptive field, improves localization accuracy, and decreases the number of parameters and FLOPs. The RepNCSPELAN4 network is introduced to optimize the YOLOv8′s neck, providing a supervisory mechanism and enhancing versatility. Experimental results demonstrate that the proposed algorithm achieves a mean average precision of 84.3%, representing a 1.05% improvement over the baseline algorithm. The accuracy rate increased by 8%, showing more powerful recognition capabilities for small-object counting like Pinus tabuliformis seeds.