In response to the unsatisfactory performance of YOLOv8n in detecting pests under complex backgrounds and dense small-object scenarios, as well as its relatively slow convergence during training, this paper proposes an improved network model named CR_ULK_YOLO (CreToNext_SPPF_UnirepLK_YOLO). Based on the YOLOv8n framework, the proposed model integrates the concept of combining large-kernel and small-kernel convolutions. A novel SPPF_UnirepLK spatial pyramid network with ultra-large kernels is designed to enhance the abstraction level of spatial features. In the neck network, a multi-branch feature fusion structure, CreToNext, is introduced to optimize both the training efficiency and inference capability of the model. Furthermore, the Shape-IoU (SIoU) loss function is adopted to accelerate model convergence and improve recognition accuracy. A customized dataset named IP9-AC, containing nine categories of agricultural pests, was constructed based on the IP102 dataset. Experimental results show that the improved CR_ULK_YOLO model achieves a mean Average Precision (mAP_50–95) of 73.4%, a recall rate of 88.4%, and a precision of 94.6%, which are 4.2%, 3.6%, and 3.4% higher, respectively, than those of YOLOv8n. Comparative experiments with other mainstream detection models demonstrate that CR_ULK_YOLO exhibits superior interpretability and practical detection performance, providing valuable insights for accurate pest identification in complex agricultural environments.