Double-branch network remote sensing image building semantic segmentation method fusing rich scale features
A remote sensing image and semantic segmentation technology, applied in neural learning methods, biological neural network models, instruments, etc., can solve the problems of incomplete segmentation and low utilization of shallow spatial features, achieve good feature fusion and ensure efficient utilization.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0045] 1. Refer to Figure 1 to Figure 9 As shown in the figure, a dual-branch network remote sensing image architectural semantic segmentation method fused with rich scale features is characterized in that, it includes the following steps:
[0046] Step1: The deep semantic path extracts different levels of building semantic features based on ResNet50 with mixed atrous convolution;
[0047] Step2: Process the extracted deep semantic features through the spatial pyramid to obtain multi-scale information in the deep features;
[0048] Step3: The shallow spatial path adopts a small downsampling multiple to maintain the resolution of the image, and mainly uses the Res2Net module and the rich-scale feature extraction module to obtain accurate image spatial information;
[0049] Step4: Feature fusion module, which adaptively fuses deep features and shallow features, and finally obtains effective segmentation results.
[0050] 1.1 Deep Semantic Path Structure
[0051] The main fun...
Embodiment 2
[0082] 2.1 Experimental data and platform
[0083] In order to fully verify the effectiveness of the algorithm in this paper, building extraction experiments are carried out using the WHU aerial dataset and Massachusetts building dataset with different spatial resolutions and large differences in building characteristics. The two datasets are described below:
[0084] (1) WHU aviation dataset. The WHU dataset is provided by the New Zealand Land Information Service (https: / / data.linz.govt.nz). The original image has a spatial resolution of 0.075m and an image size of 15354 pixels × 32507 pixels. In the experiment, the dataset is divided into 512 pixels ×512 pixel sub-image, and divided into training set, validation set, and test set. Among them, the number of images in the training set is 1330, and the number of images in the validation set and test set is 70 and 427, respectively.
[0085] (2) Massachusetts building dataset. The dataset was established by Minh, with a surfac...
Embodiment 3
[0124] In this paper, RF-DPNet is proposed to improve the problems of low utilization rate of high-resolution features in the automatic extraction of buildings from remote sensing images, poor edge segmentation effect of buildings, and unclear boundaries. This paper builds a deep semantic path and a shallow spatial path to learn from shallow and deep cross-learning to richer spatial features and global features, and then uses adaptive feature fusion to weight and constrain the shallow features rich in spatial texture information to compensate for The defects of edge and boundary information extraction in deep semantic features are eliminated, and spatial information irrelevant to building features is filtered out. The experimental results on the WHU and Massachusetts datasets show that the method proposed in this paper has higher mIoU and mPA than Deeplabv3+, BiSeNet, UNet, DenseASPP and other methods, and can accurately extract edge information. All objects can achieve better...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More