Attention-based CNN Model for Semantic Segmentation of Remote Sensing Images

The application of computer vision techniques to image segmentation tasks in remote sensing images has been steadily increasing in importance. This process partitions an image into non-overlapping segments with semantic significance. There are still problems such as insufficient segmentation accuracy, loss of small targets, segmentation gaps, and loss of information. Pixel-level classification can be performed using semantic image segmentation, which allocates a class to every pixel in an image. However, segmenting targets into multiple categories remains a challenge. The proposed system uses Convolutional Block Attention Module (CBAM) with UNet (U-shaped network) to improve the feature extraction capability of convolution layer from input image at downsampling. Downsampling and upsampling networks use self-attention layers that provide an efficient way to capture dependencies in an image. Semantic representation of feature map is reconstructed with self-attention. The feature map is decoded by upsampling to obtain the desired segmentation result. The proposed system uses channel attention in first and third upsampling stages and spatial attention in second, fourth and fifth upsampling stages at different feature fusion to improve performance across categories. Experiments are conducted on WHDLD dataset, and the results show a high IoU score for segmentation such as baresoil, building, road, and water. Mean IoU is 61.27, and 83.81% of accuracy is obtained.