摘要
针对移动终端等低算力设备,提出了一种应用于自然彩色图像的轻量化上颌前牙分割算法,选用了在自建数据集上表现较好的BiSeNet-ResNet18作为基准模型,然后对模型进行了改进。实验结果显示,改进后的BiSeNet在自建数据集上的MIoU达到了90.7%,较BiSeNet提升了1.2个百分点。在轻量化和实时性方面,模型大小只有BiSeNet的1/8,并且每秒帧数达到了63.1。该项研究成果有望在牙科医学和医疗影像学领域得到应用,推动牙齿分割、微笑美学自动分析等技术的发展。
A segmentation algorithm is proposed for lightweight maxillary anterior tooth segmentation on low‑computing power devices,such as mobile terminals.The algorithm is applied to natural color images.The benchmark model chosen for comparison is BiseNet‑ResNet18,which performs better on the self‑built dataset.The model is then further improved.The experimental results show that the mean intersection over union(MIoU)of the improved BiseNet on the self‑built dataset reaches 90.7%,which represents a 1.2 percentage improvement over BiseNet.In terms of lightweight and real‑time performance,the model size is only 1/8 of BiseNet,and it achieves a frame rate of 63.1 frames per second.The research results are expected to be applied in the fields of dental medicine and medical imaging,promoting the development of technologies such as tooth segmentation and automated analysis of smile aesthetics.
作者
吴杰
陶青川
Wu Jie;Tao Qingchuan(College of Electronics and Information Engineering,Sichuan University,Chengdu 610000,China)
出处
《现代计算机》
2023年第24期33-39,共7页
Modern Computer
