High Resolution Remote Sensing Semantic Segmentation Using Bayesian of Hyperparameters and Improved U-net

This paper proposes an improved method based on the U-net network, which aims to improve the semantic segmentation performance of high-resolution remote-sensing images, and accelerate the process of finding the optimal learning model. This model combines pyramid pooling with the U-net model to achieve feature extraction at multiple scales, and the atrous convolution is added to enlarge the receptive field, so that each result output contains a larger range of features than the original output, and the global information of remote-sensing images can be obtained. Simultaneously incorporating Bayesian optimization to optimize the model, and find the optimal parameters. The results show that compared with other commonly used semantic segmentation models, the proposed method has best segmentation accuracy and shorter training time.