Multiscale Image Segmentation using Cascade MultiResUNet Proposal Masking Convolutional Network

Segmentation of aerial view images is useful for clear identification of the objects such as detection of structures, highways, vehicles, vegetation, disaster relief, and forms of land cover. It is also useful for tracking sustainable growth, including urban resource planning, climate change, and agriculture research. By learning to use aerial imagery to automatically generate building images, segmentation techniques are becoming crucial to scaling the mapping attempts. The latest research approaches suggest that Neural Network Architectures have significance in overcoming the limitations caused by traditional segmentation techniques. However, it was still uncertain if the segmentation model was properly implemented owing to the main problem of large size and high resolution in pixel-level image segmentation. In this research work, using the Cascaded MultiResUnet Proposal Masking Convolutional Network (CMPMCN) method, segmentation of the multiscale aerial imagery is presented. The Cascaded dilated MultiResUNet method is used for identifying and achieving better objects extraction in aerial images, and also the proposal masking method is used to train for generating masks of the region at every feature map. The approach has achieved 69.2% IoU score with 980M total parameter quantity, and 3*12H training duration. The results are compared with literature methods and proven to be better.