Improved U-Net based on ResNet and SE-Net with dual attention mechanism for glottis semantic segmentation

Abstract

In previous tasks of glottis image segmentation, the position attention mechanism was rarely incorporated, neglecting the detailed information in glottis position detection. Aiming to improve the U-Net architecture, this study introduces the dual attention mechanism based on the squeeze and excitation (SE)-Net model. This mechanism can integrate traditional channel attention with position attention mechanisms to effectively adjust the weights of crucial features and significance of positions. Replacing the weight adjustment mechanism in SE-Net with the dual attention mechanism creates a broader perspective, enhancing the sensitivity to important features in the model. Furthermore, based on the characteristics of SE-Net, the skip-connection feature of U-Net can still be retained. The architecture proposed in this paper further replaces the convolutional layers in the U-Net encoder with the bottleneck to preserve the information on the features without significantly increasing the amount of computation. In addition, the decoder is replaced with residual blocks to reduce overfitting. The results of the experiment showed that models with retained features demonstrate better accuracy while reducing overfitting. The proposed model achieved positive results in predicting the scores on the public benchmark for automatic glottis segmentation (BAGLS) dataset.