Image Semantic Segmentation Model based on AsppUNet

In this paper, we propose AsppUNet, an image semantic segmentation model based on the Atrous Spatial Pyramid Pooling(ASPP) module, to address the issue that smaller objects are prone to being overlooked during the segmentation process. Instead of using the standard pooling layers in the encoder of UNet, our model adopts a series of atrous convolution layers with progressively increasing dilation rates to reduce feature loss caused by traditional pooling operations. The ASPP module is constructed by cascading atrous convolution layers with different dilation rates, and is integrated into the decoder of UNet to aggregate multi-scale feature maps and capture multi-level contextual information. Experimental results demonstrate that AsppUNet achieves superior segmentation performance on objects of various sizes. It improves the mIoU for objects at different scales on the CamVid dataset, and effectively enhances the overall segmentation accuracy.