Image semantic segmentation model based on CBAMUNet

Mainstream image semantic segmentation networks can only extract local features of an image. Their receptive field is limited to the range of convolutional kernel size, which impacts the distinguishability of extracted features by the encoder. This issue leads to inaccurate segmentation, loss of small targets, and other related problems. This paper presents a solution to the aforementioned challenges by proposing the use of a CBAMUNet image semantic segmentation model. The model combines CBAM attention module and UNet En-decoder network. Firstly, the UNet network has been enhanced to eliminate the need for cropping of input images to reduce feature loss. Secondly, the CBAM module has been integrated into its skip connection to enable the UNet network to extract more efficient features and enhance the segmentation effect. And the model uses a new loss function. Data augmentation is also used to improve the generalization ability of the model. Through the comparison experiments with UNet, it is found that CBAMUNet is able to significantly improve the segmentation effect of images, and increase the pixel precision(PA) and the mean intersection over union(mIoU) of the semantic segmentation of images.