Two-stage remote sensing target detection method based on target center point estimation

A target detection and center point technology, applied in neural learning methods, computing, computer components, etc., can solve problems such as damage to the accuracy of target detectors and unconsidered location distribution characteristics of remote sensing images.

Active Publication Date: 2021-09-10
WUHAN UNIV
6 Cites 4 Cited by

AI-Extracted Technical Summary

Problems solved by technology

However, in these existing algorithms, the location distribution characteristics of targets in remote sensing images have hardly been considered
Different from the distribution of targets in natural close-range images, in remote sensing images, there is almost no overlap of targets of the same category. Therefore, the strategy of generating target ...
View more

Method used

The two-stage target detection network based on target center point estimation includes attention-assisted feature extraction (features extraction), region proposal network (Centroid-Inference basedRegion Proposal Network, CI-RPN) based on center point recommendation, deformable volume Deformable convolution Layer, Intersection over Uniou (IoU) prediction and bounding box regression, a total of 4 parts. The basic network of the feature extraction part can use the currently popular residual network (ResidualNetwork, ResNet), stacked hourglass network (HourglassNet), deep layer aggregation network (Deep LayerAggregation, DLA), e...
View more

Abstract

The invention relates to a two-stage remote sensing target detection method based on target center point estimation. A sample library is constructed by using a remote sensing image and an annotation file of an interested category target, a two-stage remote sensing target detection network based on target center point estimation is trained, and features of the interested category target on the remote sensing image are learned. And target detection is performed on a new remote sensing image by using the trained network model, thereby realizing automatic category judgment and bounding box positioning of an interested target on the remote sensing image. According to the method, the regional suggestion network is constructed according to the spatial position distribution characteristics of the targets on the remote sensing image, and multi-category target candidate frames are directly generated. Compared with a method in which dense target candidate frames are firstly generated, then the best candidate frame is selected from the dense target candidate frames by using a non-maximum suppression method, and subsequent target identification and bounding box correction are carried out, the method of the invention has higher efficiency and precision, and is more suitable for target detection tasks of remote sensing images.

Application Domain

Character and pattern recognitionNeural architectures +1

Technology Topic

Computer visionNetwork model +2

Image

  • Two-stage remote sensing target detection method based on target center point estimation
  • Two-stage remote sensing target detection method based on target center point estimation
  • Two-stage remote sensing target detection method based on target center point estimation

Examples

  • Experimental program(1)

Example Embodiment

[0042] Embodiment:
[0043] First, in accordance with the method of the present invention, two-stage target detection network based on the target center point estimate is constructed, and CENTROID-INCENCE BASED Region CONVOLUTION NEURAL NETWORAL NETWORK, CI-RCNN) is constructed. Then obtain training sample data and use the sample data training network model. The sample data used in the examples is Ningbo Electric Tower remote sensing image data, including 228 training images and 76 test images, the image size of about 6000 × 6000 pixels, all of which have been artificial inspections. We use the original image of the training set with 512 pixels as steps, cut into video blocks of 1024 × 1024 pixels, input into the network model for iterative training until the model converges the optimal weight file. After the model training is completed, the test remote sensing image is input to the network model, and the target detection can be obtained, you can get the boundary containers of the image on the image on the image.
[0044] In order to verify the effectiveness and advancement of the method of the invention, we will compare the proposed method with other latest target detection algorithms. Includes FASTER R-CNN, PANET, YOLOV4, and CENTERNET, YOLOV4, and CENTERNET, YOLOV4, and CENTERNET, YOLOV4, and CENTERNET target detection algorithms that extends in various target detection tasks. All methods are developed using the same training data on the same hardware environment (a personal computer with NVIDIA GeForce GTX 1080Ti GPU, Intel i5-8400 CPU, Windows operating system). The prediction results of all methods were quantitatively evaluated in accordance with the COCO evaluation measure, and recorded in Table 1. From the mean average precision mean (MAP) main indicator of Table 1, the method of the present invention is superior to the other advanced target detection methods. AP 75 On the indicator, (with the measurement result between the detection result and the real target boundary box is greater than or equal to 75% as the threshold, the instance of statistical correct detection) Our method has a greater advantage over other methods. In contrast to these existing methods, it demonstrates that the method of the present invention has better robustness and can obtain more accurate target boundary frame identification and positioning results. Therefore, the method of the present invention has a good engineering practical value.
[0045] Table 1 Comparison of the method of the method of the present invention and other advanced target detection method
[0046]
[0047]

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products