{"title":"基于 YOLO 实例分割的杂草顶端分生组织定位算法","authors":"Daode Zhang, Rui Lu, Zhe Guo, Zhiyong Yang, Siqi Wang, Xinyu Hu","doi":"10.3390/agronomy14092121","DOIUrl":null,"url":null,"abstract":"Laser technology can be used to control weeds by irradiating the apical meristematic tissue (AMT) of weeds when they are still seedlings. Two factors are necessary for the successful large-scale implementation of this technique: the ability to accurately identify the apical meristematic tissue and the effectiveness of the localization algorithm used in the process. Based on this, this study proposes a lightweight weed AMT localization algorithm based on YOLO (look only once) instance segmentation. The YOLOv8n-seg network undergoes a lightweight design enhancement by integrating the FasterNet lightweight network as its backbone, resulting in the F-YOLOv8n-seg model. This modification effectively reduces the number of parameters and computational demands during the convolution process, thereby achieving a more efficient model. Subsequently, F-YOLOv8n-seg is combined with the connected domain analysis algorithm (CDA), yielding the F-YOLOv8n-seg-CDA model. This integration enables the precise localization of the AMT of weeds by calculating the center-of-mass coordinates of the connected domains. The experimental results indicate that the optimized model significantly outperforms the original model; the optimized model reduces floating-point computations by 26.7% and the model size by 38.2%. In particular, the floating-point calculation is decreased to 8.9 GFLOPs, and the model size is lowered to 4.2 MB. Comparing this improved model against YOLOv5s-seg and YOLOv10n-seg, it is lighter. Furthermore, it exhibits exceptional segmentation accuracy, with a 97.2% accuracy rate. Experimental tests conducted on five different weed species demonstrated that F-YOLOv8n-seg-CDA exhibits strong generalization capabilities. The combined accuracy of the algorithm for detecting these weeds was 81%. Notably, dicotyledonous weeds were detected with up to 94%. Additionally, the algorithm achieved an average inference speed of 82.9 frames per second. These results indicate that the algorithm is suitable for the real-time detection of apical meristematic tissues across multiple weed species. Furthermore, the experimental results demonstrated the impact of distinctive variations in weed morphology on identifying the location of the AMT of weeds. It was discovered that dicotyledonous and monocotyledonous weeds differed significantly in terms of the detection effect, with dicotyledonous weeds having significantly higher detection accuracy than monocotyledonous weeds. This discovery can offer novel insights and avenues for future investigation into the identification and location of the AMT of weeds.","PeriodicalId":7601,"journal":{"name":"Agronomy","volume":"48 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Algorithm for Locating Apical Meristematic Tissue of Weeds Based on YOLO Instance Segmentation\",\"authors\":\"Daode Zhang, Rui Lu, Zhe Guo, Zhiyong Yang, Siqi Wang, Xinyu Hu\",\"doi\":\"10.3390/agronomy14092121\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Laser technology can be used to control weeds by irradiating the apical meristematic tissue (AMT) of weeds when they are still seedlings. Two factors are necessary for the successful large-scale implementation of this technique: the ability to accurately identify the apical meristematic tissue and the effectiveness of the localization algorithm used in the process. Based on this, this study proposes a lightweight weed AMT localization algorithm based on YOLO (look only once) instance segmentation. The YOLOv8n-seg network undergoes a lightweight design enhancement by integrating the FasterNet lightweight network as its backbone, resulting in the F-YOLOv8n-seg model. This modification effectively reduces the number of parameters and computational demands during the convolution process, thereby achieving a more efficient model. Subsequently, F-YOLOv8n-seg is combined with the connected domain analysis algorithm (CDA), yielding the F-YOLOv8n-seg-CDA model. This integration enables the precise localization of the AMT of weeds by calculating the center-of-mass coordinates of the connected domains. The experimental results indicate that the optimized model significantly outperforms the original model; the optimized model reduces floating-point computations by 26.7% and the model size by 38.2%. In particular, the floating-point calculation is decreased to 8.9 GFLOPs, and the model size is lowered to 4.2 MB. Comparing this improved model against YOLOv5s-seg and YOLOv10n-seg, it is lighter. Furthermore, it exhibits exceptional segmentation accuracy, with a 97.2% accuracy rate. Experimental tests conducted on five different weed species demonstrated that F-YOLOv8n-seg-CDA exhibits strong generalization capabilities. The combined accuracy of the algorithm for detecting these weeds was 81%. Notably, dicotyledonous weeds were detected with up to 94%. Additionally, the algorithm achieved an average inference speed of 82.9 frames per second. These results indicate that the algorithm is suitable for the real-time detection of apical meristematic tissues across multiple weed species. Furthermore, the experimental results demonstrated the impact of distinctive variations in weed morphology on identifying the location of the AMT of weeds. It was discovered that dicotyledonous and monocotyledonous weeds differed significantly in terms of the detection effect, with dicotyledonous weeds having significantly higher detection accuracy than monocotyledonous weeds. This discovery can offer novel insights and avenues for future investigation into the identification and location of the AMT of weeds.\",\"PeriodicalId\":7601,\"journal\":{\"name\":\"Agronomy\",\"volume\":\"48 1\",\"pages\":\"\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-09-18\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Agronomy\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.3390/agronomy14092121\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Agronomy","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.3390/agronomy14092121","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Algorithm for Locating Apical Meristematic Tissue of Weeds Based on YOLO Instance Segmentation
Laser technology can be used to control weeds by irradiating the apical meristematic tissue (AMT) of weeds when they are still seedlings. Two factors are necessary for the successful large-scale implementation of this technique: the ability to accurately identify the apical meristematic tissue and the effectiveness of the localization algorithm used in the process. Based on this, this study proposes a lightweight weed AMT localization algorithm based on YOLO (look only once) instance segmentation. The YOLOv8n-seg network undergoes a lightweight design enhancement by integrating the FasterNet lightweight network as its backbone, resulting in the F-YOLOv8n-seg model. This modification effectively reduces the number of parameters and computational demands during the convolution process, thereby achieving a more efficient model. Subsequently, F-YOLOv8n-seg is combined with the connected domain analysis algorithm (CDA), yielding the F-YOLOv8n-seg-CDA model. This integration enables the precise localization of the AMT of weeds by calculating the center-of-mass coordinates of the connected domains. The experimental results indicate that the optimized model significantly outperforms the original model; the optimized model reduces floating-point computations by 26.7% and the model size by 38.2%. In particular, the floating-point calculation is decreased to 8.9 GFLOPs, and the model size is lowered to 4.2 MB. Comparing this improved model against YOLOv5s-seg and YOLOv10n-seg, it is lighter. Furthermore, it exhibits exceptional segmentation accuracy, with a 97.2% accuracy rate. Experimental tests conducted on five different weed species demonstrated that F-YOLOv8n-seg-CDA exhibits strong generalization capabilities. The combined accuracy of the algorithm for detecting these weeds was 81%. Notably, dicotyledonous weeds were detected with up to 94%. Additionally, the algorithm achieved an average inference speed of 82.9 frames per second. These results indicate that the algorithm is suitable for the real-time detection of apical meristematic tissues across multiple weed species. Furthermore, the experimental results demonstrated the impact of distinctive variations in weed morphology on identifying the location of the AMT of weeds. It was discovered that dicotyledonous and monocotyledonous weeds differed significantly in terms of the detection effect, with dicotyledonous weeds having significantly higher detection accuracy than monocotyledonous weeds. This discovery can offer novel insights and avenues for future investigation into the identification and location of the AMT of weeds.