Lightweight detection method for lotus seedpod in natural environment

Abstract

In order to solve the problems of the current target detection algorithms, such as poor discrimination of occluded targets, multiple parameters, complex networks, large amounts of computation, and not conducive to the deployment of mobile terminals, a lightweight lotus seedpod detection method based on YOLOv5s model was proposed in this study. First, the dataset was augmented by using a combination of offline and online augmentation, which improved the adaptability and robustness of the model in complex environments. Then, a lightweight Ghost convolution module was introduced to replace the original convolution, and a lightweight bidirectional feature pyramid network was designed, which could enhance the feature extraction and fusion capability of the network and reduce the amount of calculation and model size; On this basis, the combination of WIoU loss function and Mish activation function was adopted to improve the accuracy of feature extraction. Finally, the knowledge distillation training strategy was used to ensure the proposed lightweight model has the learning ability of a complex network model, improving the recall and precision of model detection. The results of the ablation study show that the proposed method effectively improves the detection performance of the YOLOv5s model for lotus seedpods. The mean average precision of the improved model was 89.7%, compared with the original YOLOv5s model increased by 2.8%, and the parameters and FLOPs were reduced by 2.36M and 7.3G, respectively. Compared with other detection algorithm models, the proposed algorithm model has the advantages of less computation, smaller model size, and higher detection precision. Therefore, the proposed improved optimization method based on the YOLOv5s model can effectively detect lotus seedpods, which provides theoretical research and technical support for intelligent picking of lotus seedpods in the actual operating environment.