Semantic segmentation for remote sensing images via dense feature extraction and companion loss neural network

Abstract

ABSTRACT Semantic segmentation models with good performance are crucial for the practical application of high-resolution remote-sensing images (RSI). Compared with nature images, in most cases the RSI dataset has the problem of unbalanced sample distribution between classes and unbalanced target size ratio. Using semantic segmentation to pixel-wise classify and identification RSI can solve this problem to some extent. At present, most semantic segmentation models based on mainstream networks solve these problems from the object scale and super-pixel perspective, whereas the accuracy still needs to be improved. To enhance the quality of the predicted feature maps, the Dense-Inception Net (D-INet) model is proposed based on the idea of DenseNet feature reuse and combined with the attention mechanism, which enables the network to maintain depth while widening the width to obtain more advanced semantic information. The connection of contextual information is strengthened by connecting RFB multi-scale modules at the shallow level, and shallow features with more spatial features are extracted for feature fusion with the decoder. To further lift the segmentation accuracy, a companion loss is introduced in the encoder, and the model is trained to have better segmentation performance for small sample objects by adaptively adjusting the loss coefficients. Experimental results show that the proposed method significantly increases the accuracy and mean Intersection Over Union (mIOU) scores.