Abstract

Underground pests are difficult to monitor because of their hidden activity, and traditional manual surveys are inefficient. Computer vision provides a promising approach for efficient detection; however, pest boundaries and textures in complex soil backgrounds can be obscured or confounded by soil particles, residual roots, shadows, and other factors, thereby reducing detection accuracy. In this study, an image dataset of grubs, cutworm larvae, and cutworm pupae was constructed, and a lightweight detection model, HFU-YOLO, was proposed using YOLO11n as the baseline. A Haar Wavelet Downsampling (HWD) module was introduced to reduce the loss of shallow edge and texture information. A Lightweight Frequency-Adaptive Dilated Convolution (LFADC) module was designed to enhance target feature representation and reduce soil background interference. A UGS-Lite auxiliary loss function was adopted to improve the localization stability of weak-boundary targets. The results showed that HFU-YOLO achieved Precision, Recall, mAP@0.5, and mAP@0.5:0.95 values of 77.6%, 78.3%, 83.5%, and 44.1% on the test set, respectively, representing improvements of 4.2, 3.2, 6.1, and 6.1 percentage points over YOLO11n. With 2.59 M parameters and 6.2 GFLOPs, HFU-YOLO provides a reference for intelligent recognition of underground pests in complex soil backgrounds.

IPC Classification

Keywords