A lightweight model based on YOLOv8n in wheat spike detection

Lightweight YOLOv8 for Wheat Head Detection

“…According to the data, more than 99.98% of the targets in the GWHD dataset have bounding-box areas smaller than onetenth of the image area. Drawing on earlier studies [13,14,56], we summarize our observations as follows:…”

“…Deep-learning capabilities have significantly enhanced agricultural production efficiency in general, and the models entail extracting features (e.g., wheat head color and texture), utilizing morphological operations to separate foreground from background, and employing image processing for statistical analysis [14]. Dammer et al [15] distinguished diseased wheat heads from the background by setting a binarization threshold and found a linear correlation between the disease degree of wheat ears and the visual detection results of the camera.…”

“…Wen et al [49] built upon the RetinaNet model and introduced a weighted bidirectional FPN with focal loss and attention modules, ultimately achieving good performance on both GWHD and Wheatgrass Spike Detection datasets. Ban et al [14] optimized YOLOv8 to reduce its model size without sacrificing accuracy. Research has shown that attention mechanisms enhance the discriminative capabilities of models and alleviate occlusion constraints [50].…”