Abstract:The application of artificial intelligence technology in litchi phenotype acquisition mainly focuses on object recognition, yield estimation, and picking localization. However, there is a notable lack of evaluation technology for assessing litchi growth quality throughout its entire fruiting stage. Aiming to integrate multi-source data indicators to perform a comprehensive assessment of litchi growth quality during the fruiting stage, thereby generating the evaluation profiles for litchi fruiting stages, based on the YOLO v7 network framework, an object recognition algorithm named LFS-YOLO was proposed. This algorithm enhanced recognition accuracy by mitigating errors and influences stemming from dynamic environmental backgrounds and by incorporating global attention mechanisms. Furthermore, the CIoU loss function was optimized through the inclusion of the angle between predicted regression vectors, which redefined and improved the angle penalty measure. This optimization reduced the overall degrees of freedom, thereby facilitating a more effective alignment of predicted bounding boxes with the nearest axis. By integrating multi-dimensional data, a quality evaluation function was established as the foundation for comprehensive evaluation. Experimental results indicated that the LFS-YOLO algorithm achieved a recognition accuracy of 89.1%, a precision of 92.3%, and a recall of 93.0%. The evaluation profiles generated for the litchi fruiting stage illustrated various indicators that influence growth quality throughout this stage, providing valuable insights for the advancement of comprehensive evaluation technologies pertaining to litchi fruiting stage.
