Article directory
2023
Improved Field-Based Soybean Seed Counting and Localization with Feature Level Considered（Plant Phenomics）
Abstract: The development of automated soybean seed counting tools will facilitate automated yield prediction prior to harvest and improve selection efficiency in breeding programs. An integrated method for enumeration and localization is ideal for subsequent analysis. Traditional object counting methods are labor-intensive, error-prone, and have low positioning accuracy. To quantify soybean seeds directly rather than sequentially, we propose a P2PNet-Soy method. We considered several strategies for tuning the architecture and subsequent post-processing to maximize model performance in seed counting and localization. First, unsupervised clustering is applied to incorporate overcounts at close locations. Second, high-level features are included in low-level features to provide more information. Third, nitrous convolutions with different grain sizes are applied to low-level features and high-level features to extract scale-invariant features to affect the variation of soybean size. Fourth, channel and spatial attention effectively separate the foreground from the background, facilitating the counting and localization of soybean seeds. Finally, the input image is added to these extracted features to improve the performance of the model. Using 24 soybean materials as experimental materials, the model was trained with field images of a single soybean plant obtained from one side, and the model was tested on the images obtained from the other side using the above strategy. By reducing the mean absolute error from 105.55 to 12.94, the superiority of P2PNet-Soybean over the original P2PNet in soybean seed counting and localization is confirmed. Furthermore, the trained model can be effectively applied to images obtained directly from the scene without background distractions.
Rice Plant Counting, Locating, and Sizing Method Based on High-Throughput UAV RGB Images
Abstract: Rice plant counts are crucial in yield estimation, growth diagnosis, and disaster loss assessment. Currently, rice counting still relies heavily on tedious and time-consuming manual operations. In order to ease the workload of rice counting, we adopted unmanned aerial vehicles (UAVs) to collect RGB images of rice fields. Based on this, we propose a novel rice plant counting, localization and sizing method (RiceNet), which consists of a feature extractor front-end and 3 feature decoder modules, namely density map estimator, plant position detector and plant size estimator. In RiceNet, the rice plant attention mechanism and the positive and negative losses aim to improve the ability to distinguish plants from background and the quality of estimated density maps. To verify the effectiveness of the method, we propose a new drone-based rice counting dataset, which contains 355 images and 257,793 artificially labeled points. Experimental results show that the average absolute error and root mean square error of the RiceNet are 8.6 and 11.2, respectively. Furthermore, we validate the performance of our method with two other popular crop datasets. On these three datasets, our method significantly outperforms the state-of-the-art methods. The results show that the network can accurately and efficiently estimate the number of rice, replacing traditional manual methods.
