Field corn borer larvae recognition based on multi-resolution and RIoUBCELoss

• To the scarcity of field corn borer larvae sample data, this study collects images of corn borer larvae in corn planting areas. All images are captured under natural lighting conditions, with backgrounds encompassing the leaves, stems, soil, and associated vegetation communities of corn plants, authentically reproducing the complex scenes of agricultural ecosystems. Ultimately, a dataset containing 7,304 images of corn borer larvae in the field is constructed. • To tackle the problem of insufficient model attention to the small regions of corn borer larvae in natural environments, the RIoUBCELoss is proposed in this study. This function increases the pixel weights of the regions and edges of corn borer larvae while reducing the pixel weights of the regions of background and noise, ensuring that the model can focus more on improving the prediction accuracy of the regions and edges of corn borer larvae during the optimization process. • The proposed RIoUBCELoss for field corn borer larvae identification constructs a multiscale feature space with strong position sensitivity by parallel processing of multi-resolution convolutional streams and fusing feature representations, and enhances the learning ability and expression ability of the model to the field corn borer larvae by CAM and MFE Block, while suppressing unnecessary background information. . The corn borer is the leading pest causing the reduction of corn yield in China, and accurately identifying corn borer larvae in the field is crucial for ensuring corn yield and quality. However, in field environments, the identification of corn borer larvae still faces numerous challenges, such as complex farmland ecological backgrounds, varying sizes of corn borer larvae, diverse morphological characteristics, and their low pixel ratio in images. To address these problems in the identification of field corn borer larvae, this paper proposes a field corn borer larvae recognition method based on multi-resolution representation and Regional weighted IoU with Binary Cross Entropy Loss (RIoUBCELoss). The method constructs a multiscale feature space with strong positional sensitivity by processing multi-resolution convolutional streams in parallel and integrating feature representations, which effectively reduces the interference of complex backgrounds. The convolutional attention mechanism (CAM) and multiscale feature enhancement block (MFE Block) are introduced to improve the learning ability and expression ability of the model to larval features. The proposed RIoUBCELoss focuses on optimizing the prediction accuracy of the regions and edges of field corn borer larvae, achieving precise segmentation of field corn borer larvae of varying sizes and morphological. Experimental results show that the proposed method achieves IoU of 82.22%, Dice score of 89.06%, and 95% Hausdorff distance (HD95) of 10.5764 on the field corn borer larvae image dataset. Compared to the second best model U-Net++, the IoU and Dice scores are improved by 4.34% (82.22% vs. 78.80%) and 3.27% (89.06% vs. 86.24%), respectively, while the HD95 is reduced by 24.8% (10.5764 vs. 14.0636). This demonstrates its reliable technical reference for pest identification in complex field environments.