Abstract:In order to improve the detection accuracy of clustered pepper fruits and realize the convenient and rapid deployment of the model to mobile terminal, a lightweight clustered pepper fruits detection method based on improved YOLOv8 is proposed. In the backbone network, GhostConv and the newly designed C2fGhost are used to replace the traditional convolution and C2f modules, which reduces the network parameters and computation volume, and improves the detection performance of the model; in order to improve the model's effect of detecting artichoke fruits in complex backgrounds, the global attention mechanism module is added in the neck network to improve the feature fusion capability of the model; and the SIoU boundary loss function is used to replace the original loss function, which improves the model's feature fusion capability. function is used to replace the original loss function, which improves the performance of network bounding box regression and the detection effect of small target fruits. The experimental results show that for the clustered pepper fruits dataset established in the natural environment, the number of parameters, the computational volume and the weight size of the improved model are 2.46×106, 10.5GFLOPS and 6.02MB, respectively, which are only 81%, 37% and 96% of the original network model, and the detection speed and the average accuracy reach 96Frame/s and 88.5%, respectively, which not only improves the The detection accuracy and average accuracy not only improve the detection accuracy, but also significantly reduce the overall volume of the model, which can provide reference and basis for the subsequent deployment and application of clustered pepper fruits picking robot mobile. |