摘要
针对骨骼图像特征提取存在的问题,基于MobileNetV3 large网络设计了一种融合纹理增强层的轻量级骨骼图像分类器。首先对骨骼图像进行旋转不变纹理增强处理,然后采用MobileNetV3 large网络构建分类器并对其进行训练,最后通过可视化标量对训练参数寻优。该分类器能够适应不同尺寸的骨骼图像,并对其纹理特征进行针对性训练,具有较强的鲁棒性,有效增强了梯度传播,达到分类智能化。实验表明,分类器在MURA数据集骨骼分类中的平均准确率达96.9%。
To solve the problems in bone image feature extraction, a lightweight improved bone image classifier classifier incorporating texture enhancement layers was designed based on MobileNetV3 large network.First, the bone images were processed with rotation-invariant texture enhancement.Then the MobileNetV3 large network was used to construct and to train the classifier.Finally, the training parameters were optimized by visualizing the scalars.The classifier can be adapted for different bone images and to carry out targeted training on the texture features, shows robustness, effectively enhances gradient propagation, and achieves intelligent classification.Experiments show that the average accuracy of the classifier in bone classification based on MURA dataset is 96.9%.
作者
郭子昇
王吉芳
GUO Zisheng;WANG Jifang(Mechanical Electrical Engineering School,Beijing Information Science&Technology University,Beijing 100192,China)
出处
《北京信息科技大学学报（自然科学版）》
2022年第1期90-95,共6页
Journal of Beijing Information Science and Technology University
