Tomato Seedling Stem and Leaf Segmentation Method Based on an Improved ResNet Architecture

Lina Zhang¹Xinying Li¹Zhiyin Yang¹Bo Yang²Shengpeng Yu¹Shuai Zhao¹Ziyi Huang¹Xingrui Zhang¹Han Yang¹Yixing Lin¹Helong Yu^1*Minglai Yang^1*

• ¹Jilin Agriculture University, Changchun, China
• ²Changchun University of Finance and Economics, ChangChun, China

The phenotypic traits of tomato plants reflect their growth status, and through the investigation of these phenotypic characteristics, can improve the tomato production.Traditional deep learning models encounter challenges such as large numbers of parameters, high model complexity, and susceptibility to overfitting when performing point cloud segmentation tasks. In response, this paper proposes a lightweight improved model based on the ResNet architecture.The network model has optimized the traditional residual block architecture by integrating bottleneck modules and downsampling techniques. Furthermore, by combining curvature features and geometric characteristics, this research have custom-designed and encapsulated specific convolutional layers, achieving improved segmentation results for tomato stem and leaf point clouds.Therefore, the model adopts adaptive average pooling technology, significantly enhancing its generalization ability and robustness. Experimental validation shows that the accuracy of the optimized network model for training reaches 95.11, representing a 3.26 increase compared to the traditional ResNet18 model. Compared to the testing time required by the traditional ResNet18 network model (5.37 seconds), the testing time utilized by this model is 4.02 seconds, which is 1.35 seconds shorter and a 25% increase in efficiency.Through detailed analysis of the segmented organs, four key phenotypic parameters, including plant height, stem diameter, leaf area, and leaf inclination angle, were successfully extracted and compared with manually measured data, verifying the feasibility and accuracy of threedimensional point cloud technology in extracting tomato plant phenotypic parameters.The research results indicate that the coefficients of determination (R²) for the parameters are 0.941, 0.752, 0.945, and 0.943, respectively, highlighting a high degree of correlation between these parameters and the corresponding variables. The root mean square errors (RMSE) are 0.506, 0.129, 0.980, and 3.619, respectively, providing a direct reflection of the error margins between measured and extracted values. Furthermore, the absolute percentage errors (APE) are 1.965%, 4.290%, 4.358%, and 5.526%, respectively, further quantifying the measurement accuracy. The proposed X-ResNet model exhibits