Enhancing Sunflower Head Segmentation and Counting Using Improved UNet

Highlights Proposing an improved UNet to address segmentation challenges of irregular sunflower heads. The improved UNet demonstrated strong robustness under intense sunlight and blurred conditions. A boundary enhancement module can improve the details and enhance the robustness. Abstract. Sunflowers are crucial for ensuring the global security of grain and oil supply. Timely and accurate acquisition of sunflower head numbers can improve crop management efficiency and serve as a vital basis for yield estimations. Semantic segmentation combined with image processing techniques can effectively recognize and count crops. However, accurate segmentation of individual sunflower heads remains challenging given the varying sizes and irregular shapes owing to partial occlusion. To enhance the accuracy of boundary information processing in sunflower heads and mitigate omissions, this study employed VGG16 as the UNet encoder and incorporated a boundary enhancement module in the decoder to capture detailed boundary information and improve the model’s feature learning capability and robustness. Subsequently, image processing methods were employed to segment and count adhesively connected sunflower heads. The improved UNet model's segmentation results demonstrated a Recall mean of 73.70% (95% confidence intervals (CI), 69.35%-78.06%) and a mean pixel accuracy (MPA) mean of 83.46% (95% CI, 81.61%-85.31%). It surpassed UNet, DeepLab V3+, and Convolutional Block Attention Module (CBAM)-HrNet, showing varying degrees of improvement in mean Recall and MPA: 13.04%, 1.04%, and 12.07% for Recall and 5.66%, 2.71%, and 13.58% for MPA, respectively. Furthermore, counting combining the improved UNet with image processing techniques exhibited the highest coefficient of determination (R 2 ) and the lowest root mean square error (RMSE), mean absolute error (MAE), and relative error (RE) of 0.91, 15.57, 11.86, and 0.0431, respectively, indicating the effective recognition and enumeration of sunflower heads. Keywords: Boundary enhancement, Counting, Image processing, Sunflower heads, UNet.