DLW-YOLO: an efficient algorithm for small target detection

Facing dataset quality problems, small target detection challenges, and computational resource constraints in the field of pest and disease detection, we propose a comprehensive solution. First, high-quality data support for model training is provided by constructing specialized datasets to overcome the problems of inaccurate labels, lack of sample diversity, and insufficient coverage of small target instances in publicly available datasets. Second, for the small target detection challenge, the model introduces spatial pyramid pooling efficient local aggregation network (SPPELAN) and dimension aware selective integration (DASI) techniques to significantly improve the model's ability to capture and fuse small target features, especially the detection accuracy in the complex background of farmland. Finally, based on the vision transformer via token aggregation-GSconv cross stage partial (VoV-GSCSP) framework, the light weight (LW) attention structure is designed to realize the lightweight and high efficiency of the model, to ensure the excellent performance of the model under the condition of limited computational resources, and to provide technological advantages for wheat pest and disease detection. The experiments show that the model achieves an mAP of 90.71% on the self-constructed dataset, which is a 9.69% improvement over the pre-improvement, compared with EfficientNetv2, FasterRCNN + MobileNetv2, and YOLOv5s with improvements of 14.48%, 12.59%, and 10.34%, respectively. In addition, the model has excellent generalization ability and is suitable for various crop detection, especially in multi-category and small target scenarios, and the lightweight design is convenient for deploying mobile terminals, which provides a new solution for intelligent detection of agricultural diseases.