Research On Semantic Segmentation Algorithm Of Remote Sensing Image Based On Multi-task Learning

Semantic segmentation using high-resolution remote sensing images is a challenging and key technology.However,due to the characteristics of different acquisition conditions,a large amount of information and many details of high-resolution remote sensing images,the realization of local location and classification algorithm is facing unique scientific challenges.In the past few years,multi-task learning（MTL）has been widely used in a single model to solve the problems of multiple businesses.MTL enables each task to achieve high performance and greatly reduces computational resource overhead,which also provides a new development opportunity for remote sensing image semantic segmentation.At present,many remote sensing images semantic segmentation methods using deep learning have achieved some results,but the problem of inaccurate boundary classification results has not been solved.At the same time,the problems of large network size and high resource consumption still hinder the development of remote sensing image semantic segmentation.Using the advantages of multi task learning mode,this paper mainly studies the semantic segmentation algorithm of remote sensing images based on multi task learning.The main work of this paper includes:（1）A multi-task learning model EANet based on semantic segmentation and edge detection is proposed.Aiming at the problems of unclear object edge recognition and inaccurate contour in remote sensing image semantic segmentation,edge detection is integrated into the semantic segmentation network,which further enables the segmentation network to obtain more low-level detail information and help the network better identify the object contour.Our method is very simple and can be extended to other remote sensing image object extraction tasks.（2）A joint learning framework Col Net based on image super-resolution reconstruction and semantic segmentation is proposed.Given the large and slow speed of the general semantic segmentation network model,combined with the image super-resolution reconstruction task,we reduce the size of the feature map of the image in the network flow to increase the speed of network reasoning and reduce the size of the network.This method can obtain the results of high-resolution semantic segmentation and super-resolution reconstruction by taking the image with low resolution as the input when the acquisition of high-resolution data is inconvenient or the computing resources are limited.（3）A dynamic weight method is proposed to optimize the proportion of different task optimization objectives in multi task learning.Because of the cumbersome manual setting of weights and high adjustment cost of optimization objectives in multitasking learning,combined with the idea of adaptive gradient method,we designed the adaptive dynamic weights for different main tasks in multitasking learning.We have conducted a large number of contrast and ablation experiments.Compared with not using dynamic weight,our method has achieved better results in all experiments.