An Improved UNet Model for Identifying the Voids of Hollow Fiber Membrane Images

As one of the most effective methods for mixed gas separation, the gas separation properties of Polyimide (PI) hollow fiber membranes are closely related to their void structure. However, identifying the edges of the voids can be challenging, and the varying depths of the voids cause differences in image brightness, which complicates the segmentation and recognition process. To address this issue, we propose a void segmentation method based on an improved UNet model. In the network, an attention gate (AG) and a Squeeze-and-Excitation (SE) block are employed in the skip connections to enhance feature extraction and utilization. Additionally, the residual convolution module replaces the basic convolutional structure, improving the model's ability to learn spatial features. Experimental results demonstrate that the proposed method exhibits superior awareness of shallow voids while maintaining high segmentation accuracy in void segmentation.