Rapid optical remote sensing target identification method based on depth feature recombination

Abstract

The invention belongs to the technical field of deep learning remote sensing target recognition, and particularly relates to a rapid optical remote sensing target recognition method based on depth feature recombination, which can improve the calculation speed of a target. The method comprises the following steps: respectively establishing 50 layers of ResNets network architectures and 101 layers of ResNets network architectures from bottom to top as a basis for constructing a feature pyramid network, carrying out preliminary feature extraction on a remote sensing image, and extracting four features C2, C3, C4 and C5 with different scales; and the obtained four features are mutually superposed through convolutional networks of paths from top to bottom to obtain new features M2, M3, M4 and M5 for eliminating aliasing effects among different layers. The obtained M5 feature map is doubled to obtain a new feature P5, and the feature P6 is obtained by performing 3*3 on the P5 and carrying out convolution with the step length of 2; the feature P7 is obtained by performing ReLU activation function on the feature P6, and then performing 3*3 and carrying out convolution with the step lengthof 2. The method not only has the speed advantage of a single-stage test model, but also has the calculation accuracy of a double-stage test model.

Description

technical field [0001] The invention belongs to the technical field of deep learning remote sensing target recognition, and in particular relates to a fast optical remote sensing target recognition method based on deep feature recombination to improve the calculation speed of the target. Background technique [0002] Optical remote sensing images refer to the image data acquired by sensors in the visible light and some infrared bands. Although they are easily affected by factors such as light and clouds, they are intuitive and easy to understand. The spatial resolution is usually relatively high. Under the conditions of light and clear weather, the image content is rich. , the target structure features are obvious, which is convenient for target classification and recognition. [0003] The purpose of optical remote sensing target recognition is to judge whether there is a target in the remote sensing image, and to detect, segment, feature extract and classify it. Currently,...

AI Technical Summary

Problems solved by technology

[0005] The current target recognition methods in the field of remote sensing are based on the migration of conventional target recognition methods, but they have not been well optimized for the problems of small targets and concentrated targets in remote sensing data.

At the same time, in order to obtain higher accuracy, the current methods all adopt a two-stage detection strategy, that is, first obtain the region of interest through a network, and then filter and identify the region of interest through a network. The disadvantage of this is that the parameters vary. Many, while making the forward propagation speed of the model slow

Images