U-Net-Based Semantic Segmentation and Concave Point Detection for Adhesive Droplets on Water-Sensitive Paper

Accurate quantification of droplet deposition on water-sensitive paper (WSP) is essential for evaluating the performance of spray applications in crop protection. Traditional computer vision methods, such as the Otsu algorithm, often suffer from over-segmentation and reduced robustness under varying droplet coverage conditions. Convolutional neural network based semantic segmentation models offer improved performance; however, the challenge of separating adhesive droplets remains unresolved. In this study, four classic semantic segmentation models, U-Net, SegNet, PSPNet, and DeepLabv3+, were evaluated on a dedicated WSP dataset. Results showed that the U-Net achieved the best overall performance, delivering consistently high accuracy across different coverage levels. Building upon this foundation, a multi-task learning model was developed by extending the U-Net with an additional output head for concave point detection, enabling simultaneous droplet segmentation and adhesive droplet splitting. Experimental evaluation demonstrated that the proposed model preserved the segmentation accuracy of the U-Net, while also detecting a substantial portion of concave points with relatively low false detections. These findings highlight the potential of multi-task learning for integrated WSP analysis and provide a promising step toward more accurate and automated spray deposition assessment.