MCUNet: Multidimensional cognition UNet for multi-class maize pest image segmentation

Corn leaf pest is one of the leading causes of poor quality and low yield of corn. In order to quickly and accurately identify the types and forms of corn leaf pests, this paper proposes a multi-classification corn pest segmentation network, Multidimensional Cognition UNet (MCUNet). This network can classify different pests and accurately segment the details of pest morphology. Firstly, a Nested Multi-scale Extraction (NMSE) module is designed to extract deeper dense multi-scale spatial information, enabling the network to use detailed contour information. Secondly, Multilayer Perceived Self Attention (MLPSA) is proposed to complement the long-range dependencies of feature vectors, allowing the network to better distinguish the apparent differences between target and background regions. Finally, for the multiclassification task, Squeeze-and-Excitation Networks were added to MLPSA to enhance the network’s ability to discriminate between different pests. The experimental data comes from a self-built dataset of corn pests. The experimental results show that the proposed network outperforms the current mainstream neural networks’ classification and segmentation performance.