Wheat ears detection study based on YOLOv7

Object detection is a crucial research area in computer vision. Before deep learning, methods relied on handcrafted features like Sobel, Haar, and HOG, which had limited generalization capabilities. The YOLO series of single-stage algorithms improved detection accuracy and speed by leveraging deep learning. YOLOv7 further enhanced multi-scale and small object detection. This paper introduces the SE attention mechanism and DCNV deformable convolutions into the YOLOv7 model to improve small object recognition. Experimental results show that the improved YOLOv7 increases average precision by 4.52% on the GWHD2021 wheat ear dataset, with better confidence and smaller bounding boxes.