基于改进Faster RCNN的舌部多纹理检测-《中国医学物理学杂志》

文章信息/Info

Author(s):: JI Changpeng; YANG Menghan; DAI Wei; School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China

Keywords:: Keywords: tongue multi-texture Faster RCNN deep learning deformable convolution attention mechanism transfer learning multi-target feature recognition

摘要:: 为更加高效、并行地实现舌部多纹理的识别，提出一种基于Faster RCNN的改进舌部多纹理检测方法。首先，使用可变形卷积重塑主干提取网络中的卷积层并进行可变形池化，通过实际情况调整自身形状提取目标特征，以降低漏检率；其次，引入注意力机制scSE，通过增强有意义特征提高纹理表达能力；在目标相互掺杂且目标尺度差异较大的背景下，使用加权双向特征金字塔网络进行特征融合，以提升目标检测的准确率，最后进行迁移学习。实验结果显示该方法使所有类别平均精度达到0.935，较原始Faster RCNN模型提高了0.457，说明改进后的模型能有效解决目标掺杂和多尺度差异问题，具有较高的检测精度。

Abstract:: Abstract: A Faster RCNN based method is proposed to realize the recognition of tongue multi-texture more efficiently and in parallel. Deformable convolution is used to reshape the convolutional layer in the backbone extraction network and perform deformable pooling, in which the region shape can be adjusted according to the actual condition to extract target features, thereby reducing the missed detection rate. Then, the attention mechanism scSE is introduced to improve the texture expression ability by enhancing meaningful features. In the context of multi-target hybrid and large multi-scale differences, a weighted bidirectional feature pyramid network is used for feature fusion to improve the accuracy of target detection, and finally transfer learning is conducted. The experimental results show that the method achieves an average accuracy of 0.935 for all categories, which is 0.457 higher than the original Faster RCNN model, indicating that the improved model can effectively solve the problems of multi-target hybrid and multi-scale differences, and has high detection accuracy.