Research On Semantic Segmentation Algorithm Of Remote Sensing Images Based On Deep Learning

Semantic segmentation is a pixel-level classification task that aims to assign a specific category to every pixel in an image.As a vital part of earth monitoring technology,remote sensing produces images with large size and abundant content.Semantic segmentation of remote sensing images is of great significance in urban planning,resource exploration,military reconnaissance and other fields.However,due to the complexity of remote sensing images content and optical imaging conditions,remote sensing images with sub-meter spatial resolution is up against a lot of difficulties in semantic segmentation,and it is usually arduous to obtain fine-grained and accurate semantic segmentation results.This paper conducts research on the semantic segmentation algorithm of remote sensing images based on deep learning.The main research innovations include the following two aspects:(1)Because of the long observation distance of spaceborne cameras,tiny objects in highresolution remote sensing images usually occupy only a few pixels.These tiny objects are often surrounded by the dominant categories in remote sensing images,resulting in highly confusing category boundaries.In order to solve the above bottleneck problem,This paper proposes a feature-guided semantic segmentation network for remote sensing images based on encoder-decoder structure.This paper divides the features from the encoder into two parts:edge features related to object contours and body features related to object interiors,and explicitly extract edge features and body features respectively from shallow layers and deep layers.In the decoder,a feature guided module is designed to guide semantic segmentation prediction through skip connections,in which edge features are used to enhance the edge contours of objects,and body features are used to strengthen the internal consistency of objects.In addition,the multi-task learning strategy is adopted by introducing edge detection task and auxiliary segmentation task,and the performance of the final semantic segmentation task is boosted in virtue of the correlation between different tasks.(2)There are abundant ground objects in remote sensing images,which are very different in scale,texture and geometric structure.In this paper,a dual expectation maximization attention module is proposed to fully aggregate multi-scale contextual information and adapt to a large number of objects with different scales and shapes in remote sensing images.This module utilizes the self-attention mechanism to simultaneously capture contextual information in both spatial and channel dimensions,enhancing the representation of morphological features at different scales.In addition,this paper obtains a set of low-dimensional bases by introducing an expectation maximization algorithm,and calculates the attention weight matrix on this set of bases instead of all the global pixel positions,which significantly reduces the complexity of the algorithm.In this paper,an attention mechanism is adopted to capture local and global context information and enhance the representation of objects at different scales.Since the self attention mechanism needs to calculate the global autocorrelation matrix,the algorithm has high computation cost.For higher resolution remote sensing images,this resource consumption is unbearable.Therefore,this paper proposes a dual attention mechanism.Through the expectation maximization algorithm,a set of lower dimensional bases are obtained in the spatial and channel dimensions respectively.The attention weight matrix is calculated on this set of bases instead of all global pixel positions,which reduces the complexity of the algorithm and captures the context information constraints in space and channel simultaneously.Experiments on two large-scale remote sensing images semantic segmentation datasets(Vaihingen and Potsdam)demonstrate the effectiveness and advancement of the algorithm proposed in this paper.Compared with the global attention mechanism,the parameter amount of the dual expectation-maximization attention module proposed in this paper is reduced by11%,and the number of floating point operations is reduced by 41%.Compared with existing state-of-the-art semantic segmentation methods,the proposed method achieves the best results on both datasets.On the Potsdam dataset,the overall accuracy can reach 91.24%,and the mean intersection over union can reach 86.80%.