AN EFFICIENT AND ACCURATE GHOST-OPTIMIZED LIGHTWEIGHT U-NET FOR SATELLITE IMAGE SEGMENTATION

Deepak Gupta, Om Prakash Singh, Satyasundara Mahapatra

Abstract: Identifying roads, vegetation, land, and water resources is crucial for the sustainable growth and improvement of remote areas globally. For earth monitoring, computing semantic segmentation within the limits of constrained computing conditions is both an opportunity and a challenge. This study addresses the challenge of maintaining the quality of semantically segmented images by introducing an enhanced lightweight architecture. The proposed model provides a low computational cost solution for generating automatic area segmentation for the MBRSC dataset. To achieve optimal performance, an extensive comparison of patch sizes is made to choose patches of size 256x256 from the various tiles during image preprocessing. The proposed lightweight U-Net model enhances the mathematical and structural complexity of the encoder as well as the decoder by incorporating ghost convolutions. The model’s performance is assessed using a custom loss function that uses focal, dice, and tversky loss. Ghost convolution-based U-Net is proved the best option as it demonstrated the ideal efficiency-accuracy trade-off due to outstanding computational efficiency of 0.38 GFLOPs, 1.88 MB size and maintains high segmentation quality with 87.3% test accuracy, 58.4% MIoU. The proposed architecture has achieved better efficiency ratios of 153.68 MIoU/GFLOP and 31.06 MIoU/MB.

Keywords: Ghost Convolution, U-Net, Satellite Image Segmentation, Lightweight CNN

DOI: 10.24874/PES08.02.006

Recieved:   Revised:   Accepted:   
UDC:

Reads: 2   

---