Pest Recognition and Classification Using Hybrid Quantum Convolution and Diverse Branch Block
Zhuxiang Mo, Xu Bao, Yi Li, Yun Wang, Ce Wang, Li Feng · Journal of Applied Entomology · 2025
ABSTRACT The application of deep learning (DL) holds substantial value within entomology, particularly for the identification and classification of insect pests. Because pests vary widely in size and exhibit complex behavioural patterns, existing CNN‐based pest localization algorithms struggle to meet the demands of agricultural pest management effectively. In contrast, quantum systems can exploit the high dimensionality of quantum state space to effectively encode and process highly complex features. In this context, our research focuses on developing a more precise and efficient pest detection method. We have created a new dataset named AmwayYCD, which contains 1086 images of pests from six different categories. Additionally, we utilised the publicly available IP102 dataset for model evaluation and comparison, comprising 75,222 images across 102 categories. The proposed model, named YOLOv8‐QCD, is based on an improved YOLOv8 architecture with hybrid quantum convolution and a diverse branch block (DBB). This network enhances the feature extraction process through the application of DBB techniques and the Quantum Spatial Pyramid Pooling Fast (QSPPF), whereby the incorporation of quantum convolution optimises feature transformation, augmenting the model's capacity to encapsulate multi‐scale pest attributes. Experimental results demonstrate significant improvements, with YOLOv8‐QCD achieving 98.72% accuracy on AmwayYCD (1.07% higher than YOLOv8) and 75.92% on IP102 (3.78% improvement over baseline), while maintaining superior computational efficiency (18.5 M parameters, 14.4 GFLOPS). Ablation studies confirm the complementary effects of quantum convolution and DBB modules, contributing 0.52%–2.85% accuracy gains. Statistical significance testing ( p < 0.05) validates its superiority over state‐of‐the‐art methods, including 1.23% higher accuracy than LSMAE‐based Transformer on IP102. The model achieves an inference speed of 217 FPS, underscoring the remarkable potential of integrating quantum technology with deep learning for real‐time pest detection in agricultural settings—a breakthrough that offers a new direction for advancing the field of entomology.