GhostConv+CA-YOLOv8n: a lightweight network for rice pest detection based on the aggregation of low-level features in real-world complex backgrounds

Deep learning models for rice pest detection often face performance degradation in real-world field environments due to complex backgrounds and limited computational resources. Existing approaches suffer from two critical limitations: (1) inadequate feature representation under occlusion and scale variations, and (2) excessive computational costs for edge deployment. To overcome these limitations, this paper introduces GhostConv+CA-YOLOv8n, a lightweight object detection framework was proposed, which incorporates several innovative features: GhostConv replaces standard convolutional operations with computationally efficient ghost modules in the YOLOv8n’s backbone structure, reducing parameters by 40,458 while maintaining feature richness; a Context Aggregation (CA) module is applied after the large and medium-sized feature maps were output by the YOLOv8n’s neck structure. This module enhance low-level feature representation by fusing global and local context, which is particularly effective for detecting occluded pests in complex environments; Shape-IoU, which improves bounding box regression by accounting for target morphology, and Slide Loss, which addresses class imbalance by dynamically adjusting sample weighting during training were employed. Comprehensive evaluations on the Ricepest15 dataset, GhostConv+CA-YOLOv8n achieves 89.959% precision and 82.258% recall with improvements of 3.657% and 11.59%, and the model parameter reduced 1.34%, over the YOLOv8n baseline while maintaining a high mAP (94.527% vs. 84.994% baseline). Furthermore, the model shows strong generalization, achieving a 4.49%, 5.452%, and 3.407% improvement in F1-score, precision, and recall on the IP102 benchmark. This study bridges the gap between accuracy and efficiency for in field pest detection, providing a practical solution for real-time rice monitoring in smart agriculture systems.