Fast and Accurate Density Estimation of Hybrid Rice Seedlings Using a Smartphone and an Improved YOLOv8 Model

In seedling cultivation of hybrid rice, fast estimation of seedling density is of great significance for classifying seedling cultivation. This research presents an improved YOLOv8 model for estimating seedling density at the needle leaf stage. Firstly, the auxiliary frame technology was used to address the problem of locating the detection area of seedlings. Secondly, the Standard Convolution (SConv) layers in the neck network were replaced by the Group Shuffle Convolution (GSConv) layer to lightweight the model. A dynamic head module was added to the head network to enhance the capability of the model to identify seedlings. The CIoU loss function was replaced by the EIoU loss function, enhancing the convergence speed of the model. The results showed that the improved model achieved an average precision of 96.4%; the parameters and floating-point computations (FLOPs) were 7.2 M and 2.4 G. In contrast with the original model, the parameters and FLOPs were reduced by 0.9 M and 0.6 G, and the average precision was improved by 1.9%. Compared with state-of-the-art models such as YOLOv7 et al., the improved YOLOv8 achieved preferred comprehensive performance. Finally, a fast estimation system for hybrid rice seedling density was developed using a smartphone and the improved YOLOv8. The average inference time for each image was 8.5 ms, and the average relative error of detection was 4.98%. The fast estimation system realized portable real-time detection of seedling density, providing technical support for classifying seedling cultivation of hybrid rice.