Hyper-AKYOLO: a multi-scale object detection system for farmland based on hypergraph enhancement and AKConv

Abstract Rainy weather in farmland often leads to blurred images, low contrast, and degraded edges, causing high missed detection rates of tiny and multi-scale pests and disease spots. To address these challenges, this paper proposes Hyper-AKYOLO, a multi-scale object detection system that integrates Adaptive Kernel Convolution (AKConv) and hypergraph enhancement. AKConv is embedded throughout the backbone and feature pyramid to dynamically adjust kernel sampling positions and weights, enabling adaptive perception of irregular and small lesions. A hypergraph module partitions high-level features into semantic nodes and builds three types of hyperedges—semantic similarity, spatial proximity, and texture consistency—to model complex high-order relationships. Hypergraph convolution then propagates contextual information, leveraging agricultural priors such as clustered disease spread, thereby enhancing robustness in low-quality images. Experiments on a rainy farmland dataset show that Hyper-AKYOLO achieves an AP of 86.7% with a missed detection rate of 7.0%, outperforming YOLOv8s by 7.8% AP and reducing missed detections by 10.1%. The model also demonstrates higher stability across light and heavy rain scenarios while maintaining real-time inference capability. These results confirm that Hyper-AKYOLO offers a robust and practical solution for intelligent farmland monitoring under adverse weather conditions.