MSHF-YOLO: Cotton growth detection algorithm integrated multi-semantic and high-frequency features

Abstract

Accurate monitoring of cotton growth is essential for precision agriculture. However, existing deep learning-based object detection models often underperform in complex field environments due to challenges such as occlusion and low contrast. To address these limitations, we propose MSHF-YOLO, an improved detection framework based on YOLOv8. The model incorporates a Multi-Semantic Spatial and Channel Attention (MSCA) module in the backbone to enhance feature representation. Additionally, we replace traditional upsampling and downsampling operations in the neck with DySample and Adaptive Wavelet Down (AWD) modules to preserve high-frequency information. A High-frequency boost (HB) module is further introduced in the detection head to enhance detail sensitivity. Experimental results demonstrate that MSHF-YOLO achieves [email protected] of 86.0% and [email protected] of 68.2%, outperforming the baseline by 5.5% and 3.5%, respectively, while reducing model size by 12.5%. These results highlight the model's effectiveness and potential for robust cotton growth monitoring.