Semantic Segmentation of Crop Pests and Diseases Based on Dynamic Downsampling Randomized Path Convolution

Accurate detection of crop pests is of great significance for safeguarding agricultural production and food security. Traditional pest identification methods rely on manual inspection, which is inefficient and subjective, while deep learning-based algorithms can significantly improve the accuracy and automation of detection. In this study, an improved YOLOv11 instance segmentation model, YOLOv11-DDSPConv, is proposed to optimise the feature extraction capability by introducing a dynamic downsampling stochastic path convolution module (DDSPConv).The DDSPConv module employs a dual-branching parallel processing mechanism combined with a spatial Statistical Fusion Module (SSFM) and Optimised Dynamic Convolution Module (OCC), which dynamically adjusts the parameters of convolution kernel and enhances the adaptability of the model to small targets and complex backgrounds. In addition, the Path Dropout mechanism is introduced in the module to enhance the robustness of the model by randomly dropping the auxiliary paths in the residual connections. The experiments are validated using the publicly available crop pest dataset from the Roboflow platform under the NVIDIA GeForce RTX 3060 hardware environment. The results show that the improved YOLOv11-DDSPConv model achieves 90.3% on the Mask-mAP50 metric, which is a 1.8% improvement over the baseline model YOLOv11-seg, while the number of parameters is only 6.6M, which is better than other improved versions. The ablation experiment further validates the effectiveness of the DDSPConv module, which improves the detection accuracy while maintaining the lightweight nature of the model. This study provides an efficient solution for real-time accurate detection of crop pests, and the performance can be further improved by model lightweighting and dataset optimisation in the future.