HDMS-YOLO: A Multi-scale Weed Detection Model for Complex Farmland Environments

Hua Jing^*Ruimin HeYanhua ZengQi Chen^*

• Jiangxi Agricultural University, 南昌市, China

With the continuous development of agricultural technology, automatic weed removal has enormous market potential. Nevertheless, the diversity and different sizes of weeds, along with the high visual similarity between weeds and crops in terms of shape, color, and texture, make accurate weed identification a persistent challenge. Our research focuses on weed identification and trains and evaluates the HDMS-YOLO model using the CropAndWeed dataset. Specifically, to extract more weed features, the model integrates two novel feature extraction components: the SRFD module for extracting shallow features and the DRFD module for extracting deep features. These two modules complement each other, enhancing the network's ability to extract weed features. The PC-MSFA module replaces the traditional C3K2 with partial convolution and residual connections, thereby improving the model's representation of weed features. The new IntegraDet dynamic task alignment detection head enhances the model's accuracy in localization and classification. Experimental results show that the accuracy, recall rate, and mAP values of HDMS-YOLO are 74.2%, 66.3%, and 71.2%, respectively, 2.6%, 2.1%, and 2.6% higher than those of YOLO11. Compared with other mainstream algorithms, HDMS-YOLO has the best overall performance. HDMS-YOLO exhibits outstanding overall performance, supporting farm management and intelligent weed removal robots in unmanned farms.