Abstract
House segmentation of remote sensing image based on deep learning has become the main segmentation method because it can automatically extract features. However, the accuracy of image segmentation is affected not only by the network model, but also by the loss function, but the existing loss functions, except Binary Cross Entropy, are designed to deal with imbalanced dataset, no new research on improving Binary Cross Entropy for balanced dataset, and all loss function treat each pixel in isolation, without considering the spatial correlation between pixel and its neighbor pixels. To solve this problem, a new loss function, named NeighborLoss function, is proposed. Firstly, the deep learning network is used to get the prediction results of each pixel. According to whether the prediction results of the eight neighboring pixels of each pixel are consistent with the each pixel prediction, different weights are given to each pixel. Finally, the weighted average value of cross entropy of all pixels in the batch is taken as the final loss function value. We use the main deep learning semantic segmentation networks SegNet, PSPNet, UNET ++, MUNet with both NeighborLoss and cross entropy loss respectively to extract houses on the open data set named WHU dataset for remote sensing. The results show that compared with cross entropy loss functions, the MIoU, Precision, Recall, and Accuracy of NeighborLoss function are improved. From the predicted graph, the NeighborLoss function is more accurate to extract the edge of the house, especially in the corner of the house. NeighborLoss function is a more effective loss function for remote sensing image segmentation.
Highlights
The traditional semantic segmentation of remote sensing image is to use the spectral characteristics of the image, set a threshold after transformation by some algorithm, and classify the pixels with similar values [1]–[3]
In order to overcome the situation that small samples are submerged by background, FocalLoss and other variant of Cross Entropy Loss give different weights to each pixel according to foreground or background, the average cross entropy of all pixels is calculated as the final loss value
WORK The accuracy of image segmentation is affected by the network model, and by the loss function, but existing loss function are only improved for imbalanced dataset, no new research on improving cross entropy loss for balanced dataset, and all loss function treat pixels in isolation
Summary
The traditional semantic segmentation of remote sensing image is to use the spectral characteristics of the image, set a threshold after transformation by some algorithm, and classify the pixels with similar values [1]–[3]. DISTRIBUTION-BASED LOSS Binary Cross Entropy Loss [24] is the most widely used loss function in the field of image semantic segmentation It independently evaluates the cross entropy of each pixel prediction value and label, and averages all pixels, so the pixels in the image are learned . In the dataset with imbalanced categories, the cross entropy loss function will lead to training dominated by categories with more pixels, and it is difficult to learn the features of smaller objects, reducing the effectiveness of the network. 2) In this paper, we propose a practical method which takes into account the spatial correlation of neighborhood, and it is proved by experiments that the change of super parameters has little effect on the segmentation accuracy, we can directly use fixed values to avoid the problem of parameter adjustment. 3) The loss function proposed in this paper, unlike other loss function which designed to solve the problem of imbalanced data sets, can achieve better results than Binary Cross Entropy on balanced data sets, and can replace cross entropy loss function in the future
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
