MAVM-UNet: multiscale aggregated vision MambaU-Net for field rice pest detection.

Pests in rice fields not only affect the yield and quality of rice but also cause serious ecological and environmental problems due to the heavy reliance on pesticides. Since various pests have irregular and changeable shapes, small sizes, and complex backgrounds, field rice pest detection is an essential prerequisite and challenge for the precise control of pests in the field. A multiscale aggregated vision MambaU-Net (MAVM-UNet) model for rice pest detection is constructed. The model consists of four main modules, Visual State Space (VSS), multiscale VSS (MSVSS), Channel-Aware VSS (CAVSS), and multiscale attention aggregation (MSAA), where VSS is used as the basic module for capturing context information, MSVSS is used to capture and aggregate fine-grained multiscale feature of field rice pest images, CAVSS is added into Skip connection to select the critical channel representations of the encoder and decoder, and MSAA is added in the bottleneck layer to integrate the pest features of different layers of the encoder. Combining MSAA and CAVSS can capture the low-level details and high-level semantics and dynamically adjust the contributions of features at different scales; for example, the slender legs and antennae of pests rely on fine-grained features, while the large body of pests relies on coarse-grained features. A large number of experimental results on the rice pest image subset of the IP102 dataset show that MAVM-UNet is superior to the state-of-the-art models, with