Developing a More Frequent and Faster Method for Segmentation of Urban Roads From Satellite Images

Abstract

In the recent era of computer vision research and development, the researchers still struggle in the case of an upkeep in contribution to this genre. Image segmentation is the technique of dividing a picture into useful regions and items. It may be applied to comprehending and recognizing scenes in a variety of industries, including ecology, healthcare, automation, and aerial photographs. So here we propose, an automatic method for recognizing the urban network of highways on high resolution satellite images has been developed and implemented. Then we also did several experiments regarding our satellite-based image dataset with the U-Net, U-Net with ResNet50 as an encoder, DeepLabV3+ ResNet50 as an encoder and the DeepLabV3+ ResNet101 for the comparison experimentation. Our main purpose of this research work was to apprehend between the last decade of research to our own accorded model. A comparison between the results was then introduced in our research article. In such circumstance we can say that our result was divided into various contexts of the models in order of our result convention. In case of our result estimation the best performance was acquired by the DeepLabV3+ ResNet101, the dice loss is 0.05 and the mean IoU 90%. The U-Net ResNet50 got the same IoU of 90%, but the dice loss was 0.06. Which and why we would prefer our model of the DeepLabV3+ ResNet101 to be the best in case. We prepared out dataset in order of accustom to our models structure. We hope our work would be enlightened in case of the image segmentation and computer visions massive indenture.