An unmanned aerial vehicle aerial image change detection algorithm based on semantic segmentation

Abstract

According to the unmanned aerial vehicle aerial image change detection algorithm based on semantic segmentation, an unmanned aerial vehicle is used for shooting the same place in two different time periods to collect image data, an original image is obtained, and the algorithm is characterized by comprising the following steps that S1, making a label coding set; S2, making a training set; S3, training to generate a semantic segmentation network; S4, obtaining two semantic segmentation result images; and S5, obtaining a change detection result. According to the algorithm, the semantic level characteristics of the image are fully utilized, some auxiliary training sets are constructed for training, so that the trained network can learn some generalized characteristics, and the final detectioneffect is higher than that of the traditional manual characteristic. According to the method, the final change result can judge whether the change occurs or not, the change category can be detected,16 change categories can be recognized at present, and some requirements for change detection in actual research problems are better met.

Description

technical field [0001] The invention belongs to the technical field of UAV aerial photography image processing, and in particular relates to an algorithm for detecting changes in UAV aerial photography images based on semantic segmentation. Background technique [0002] Change detection is the process of identifying state changes through multiple observations of ground objects or phenomena. At present, the change detection technology of multi-temporal images has been widely used in land survey, urban research, ecosystem detection, disaster detection and other applications. [0003] So far, there are two main types of mainstream change detection frameworks. One is the pixel-based change detection algorithm, which mainly generates a difference map by comparing two images in the same area at different times pixel by pixel, and then The generated difference map is used for image segmentation operation, and finally the change result map is obtained. Such methods are generally in...

AI Technical Summary

Problems solved by technology

However, this method is too dependent on the segmentation results, and because it is all based on object detection, the preservation of details is often not good enough.

[0004] The existing patent 201610163983.X "SAR Image Change Detection Based on Deep Learning and SIFI Features" mainly uses the scale-invariant feature change method to extract the SIFT features of the image, and then uses the SIFT features as samples to train a neural network, using traditional image algebra The feature difference map is obtained by using the method, and the neighborhood features of each pixel in the difference map are extracted as test data, which are input into the trained neural network for testing, and the final change detection results are output; the main process is to read in the image-image regression Synthesis-construct training features-reduce the feature dimension and input it into the deep neural network-calculate the log ratio difference image of two images-construct the field feature sample matrix of the log ratio difference image-detect the log ratio difference image-output detection categories; this invention mainly utilizes the stability of SIFT features for image noise, overcomes the influence of SAR image speckle noise, and improves the accuracy of SAR image change detection; although this invention uses neural networks, its input features The traditional SIFT feature is mainly used, and the semantic level of the image is not used. The semantic level of the image means that the algorithm knows the actual feature represented by the pixel and its semantic information when classifying the elements in the image. Rather than relying on people to define each category after separating categories like the traditional method, the network structure is also a simple restricted Boltzmann machine, which does not fully utilize the advantages of deep neural networks in image processing.

At the same time, the final result of the invention is still only able to identify the changed area, that is, to judge whether there has been a change, but not to identify the changed category

Images