Research on an Improved Box Plot Key Point Detection Algorithm Based on YOLOv8

Box plot are a kind of common statistical charts in biological sciences and other literatures. In view of the difficulty of automatically obtaining the retrospectively annotated original data and accurately detecting the key data points in box plot in the published literatures, this paper proposes a lightweight detection model based on the improvement of YOLOv8-pose. The model firstly uses ShuffleNetv2 instead of its backbone network to reduce the number of parameters and computation, secondly introduces SE attention mechanism in the output of the backbone network to reduce the interference of redundant information, and finally employs BiFPN module in the feature fusion network at the neck for better fusion of low-level features. In the paper, taking the box plot graph dataset constructed by the authors as an example, with the help of simulation experiments, we verify that the detection accuracy of the improved model remains basically unchanged, the amount of floating-point computation is reduced by 61.4 %, the amount of parameters is reduced by 60.1 %, and the size of the model is reduced by 54.23 % compared with the original model. The results show that the improved lightweight model can significantly reduce the amount of floating-point computation and the number of parameters while maintaining high detection accuracy, which makes it easier to deploy the model at a later stage.