Research On Remote Sensing Image Segmentation Based On Multi-receptive Field Fusion And Pseudo-label Optimizatio

With the continuous development of remote sensing observation technology,an increasing number of high-resolution remote sensing images are widely applied in various fields.High-resolution remote sensing images contain complex semantic information,therefore,remote sensing image segmentation is challenging.Based on deep learning,semantic segmentation of remote sensing images can divide complex semantic information into multiple independent regions,and assign a category to each pixel,which is an important research direction in the field of remote sensing image.Although a lot of research has been done on deep learning-based remote sensing image semantic segmentation both at home and abroad,there are still some shortcomings.For example,the texture details in remote sensing images are richer than those in natural images,and the scale of objects in the image varies greatly,which may lead to misclassification or omission.In addition,the labeling of remote sensing image segmentation datasets often requires professionals to annotate each pixel in the image one by one,and the entire process consumes a lot of manpower and financial resources,which severely limits the application of convolutional neural networks in remote sensing image segmentation.Taking into account the above problems,this article carries out in-depth research in both fully supervised and semi-supervised directions,with the main work as follows.(1)Due to the large scale variation and complex category information of targets in remote sensing images,this article proposes a deep learning network based on multi-receptive field feature fusion.The network takes residual network as the backbone structure,and inserts a multi-receptive field fusion module after the backbone network to obtain multi-scale receptive field information through rectangular convolutions with different aspect ratios.Furthermore,a new attention fusion module is proposed,which uses attention mechanism to organically combine deep semantic information content with shallow semantic information content,and this fusion strategy enhances the connection between category information.Finally,the network output end is fused with a feature refinement module to gradually restore the high-resolution image and refine the category features.The experimental results on two datasets provided by ISPRS demonstrate that the proposed network can effectively handle various targets in complex environments.(2)In fully supervised remote sensing image segmentation tasks,the cost of data annotation is high,and insufficient annotated data can degrade the performance of deep learning algorithms.To address this issue,this paper proposes a semi-supervised remote sensing image segmentation network(PTONet)based on pseudo-label optimization.This method draws on the idea of consistency regularization and uses a mean teacher network structure as the base model.First,the student network is trained with labeled data,and then the student network is jointly trained with unlabeled data,while the teacher network updates its parameters through exponential moving average.When generating pseudo-labels from unlabeled data,this paper uses information entropy to calculate the classification quality of pseudo-labels.High-quality pseudo-labels are used for unsupervised learning,while lowquality pseudo-labels are used for negative sample screening.The screened negative samples are sent to the category library for iterative updates,and finally,the entire network is supervised by comparing losses.The experimental results also demonstrate that this paper’s network can achieve better results with a small amount of annotated data.