1Graduate School of Agriculture and Life Sciences, The University of Tokyo, Tokyo, Japan.
2Institute of Crop Sciences, National Agriculture and Food Research Organization, Tsukuba, Ibaraki, Japan
3Western Region Agricultural Research Center, National Agriculture and Food Research Organization, Fukuyama, Hiroshima, Japan
4Tohoku Agricultural Research Center, National Agriculture and Food Research Organization, Morioka, Iwate, Japan
| Received 22 Sep 2022 |
Accepted 01 Feb 2023 |
Published 15 Mar 2023 |
Developing automated soybean seed counting tools will help automate yield prediction before harvesting and improving selection efficiency in breeding programs. An integrated approach for counting and localization is ideal for subsequent analysis. The traditional method of object counting is labor-intensive and error-prone and has low localization accuracy. To quantify soybean seed directly rather than sequentially, we propose a P2PNet-Soy method. Several strategies were considered to adjust the architecture and subsequent postprocessing to maximize model performance in seed counting and localization. First, unsupervised clustering was applied to merge closely located overcounts. Second, low-level features were included with high-level features to provide more information. Third, atrous convolution with different kernel sizes was applied to low- and high-level features to extract scale-invariant features to factor in soybean size variation. Fourth, channel and spatial attention effectively separated the foreground and background for easier soybean seed counting and localization. At last, the input image was added to these extracted features to improve model performance. Using 24 soybean accessions as experimental materials, we trained the model on field images of individual soybean plants obtained from one side and tested them on images obtained from the opposite side, with all the above strategies. The superiority of the proposed P2PNet-Soy in soybean seed counting and localization over the original P2PNet was confirmed by a reduction in the value of the mean absolute error, from 105.55 to 12.94. Furthermore, the trained model worked effectively on images obtained directly from the field without background interference.
