Abstract:

During the general and full−bloom phases of potatoes, flowers were observed to be particularly small and densely distributed. In real−scene potato farmland images, the features of the flowers occupied minimal pixel regions, leading to frequently missed detections and low accuracy in mainstream recognition models. To address this issue and improve flower recognition accuracy, the YOLOv5CD5−256 model was proposed. Inspired by DenseNet, a CD5−256 dense structure was integrated into YOLOv5 by incorporating DenseBlocks, and attention mechanisms were incorporated to enhance feature extraction. For validation, images captured by eight real−scene monitoring systems in Ningxia potato fields from 2019 to 2023 were used as the dataset. The proposed model was compared with YOLOv5L, YOLOv8L, YOLOv8X, YOLOv9C and YOLOv9E. The results showed that the precision (P), recall (R) and mean average precision (mAP) of YOLOv5CD5−256 on the test set reached 0.83, 0.85 and 0.82, respectively. Each of these indicators was 0.20 higher than those of YOLOv5L and 0.15−0.17 higher than those of other models. It performed the best among the six models. In the early flowering stage of potatoes, all six models had good detection capabilities. When potatoes entered the full−bloom and peak−bloom stages, the average missed− detection rate of the YOLOv5CD5−256 model was 0.20−0.23 lower than that of other models, showing obvious advantages. This indicates that the proposed model can be applied to flower detection in different stages of potatoes, including the early, full−bloom and peak-bloom stages. Notably, its detection ability for small−feature and densely distributed flowers is significantly better than that of current mainstream models, and it can be used as a recognition model for potato flowers.

Key words: Potato, Flower recognition, YOLOv5, Densenet, Attention mechanism