Intelligent annular mountain detection method based on UAED model

Abstract

The invention discloses an intelligent annular mountain detection method based on a UAED model. The method comprises the following steps: constructing a training set and a verification set based on a public data set; designing an annular mountain UAED detection model structure; designing a loss function according to the model structure; training the UAED detection model by using the training set, obtaining a series of trained UAED detection models through minimizing a loss function, and selecting the finally used UAED detection model based on the verification set; deploying the trained UAED detection model on a satellite to form a UAED detection reasoning module; an original remote sensing image is generated on the satellite and input into the UAED detection reasoning module, and the edge of the annular mountain is detected and obtained and used for assisting in subsequent landing point recognition. According to the method, on-orbit detection of the lunar surface image annular mountain is realized, and the problems that a traditional annular mountain detection method is low in recognition rate and poor in flexibility, and data samples of lunar surface annular mountain detection are missing are solved.

Description

Technical Field

[0001] The present invention relates to an intelligent crater detection method based on the UAED model, which can be used for on-orbit crater detection in lunar surface images, especially for large-scale remote sensing image processing in orbit, and belongs to the field of intelligent data processing of payloads. Background Art

[0002] During the process of the detector landing and selecting a suitable landing site, due to the complexity and particularity of the space environment, the following challenges will be faced:

[0003] (1) Small target area: Due to the complex terrain on the lunar surface, there is a large difference in the sizes of effective reference objects such as craters, and the distance between the detector and the lunar surface during landing ranges from 0 to 100 km, resulting in small targets in the images captured by the detector. The pixels of these small targets are filtered out during the process of the target recognition module extracting semantic information, making it difficult for the small targets to be captured by the target detection module with semantic information. Therefore, the detection effect of small targets is usually poor.

[0004] (2) Few sample types: Although the lunar surface environment is relatively complex, the number of its terrain types is small, and the number of sample types that can be obtained is small. Despite the huge amount of visual data that the detector can obtain, most of it is repetitive, and the proportion of valuable data is small, resulting in inapplicability to related modules that rely on large-scale data sets.

[0005] Existing intelligent crater detection methods have problems: when manually annotating the edges of images, the results annotated by different people are different, and how to effectively utilize multiple annotations is a challenge. Traditional methods adopt the way of randomly selecting one annotation or averaging multiple annotations, and the effect is not good. This patent introduces sample variance to solve this problem and better utilizes the correlation and difference between multiple annotations. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: overcoming the deficiencies of the prior art, providing an intelligent crater detection method based on the UAED model, which realizes on-orbit detection of lunar surface crater images based on the edge detection idea and adopts the method of ground training and on-orbit inference, and solves the problems of low recognition rate, poor flexibility of traditional crater detection methods, and lack of lunar surface crater detection data samples.

[0007] The technical solution of the present invention is: In the first aspect, an intelligent crater detection method based on the UAED model is provided, including:

[0008] Constructing a lunar surface crater detection data set based on a public data set, including a training set and a validation set;

[0009] Designing the structure of the crater UAED detection model;

[0010] Design a loss function according to the model structure;

[0011] Use the training set to train the UAED detection model. After obtaining a series of trained UAED detection models by minimizing the loss function, select the finally used UAED detection model based on the validation set;

[0012] Deploy the trained UAED detection model on the satellite to form a UAED detection inference module;

[0013] Generate the original remote sensing image on the satellite and input it into the UAED detection inference module to detect and obtain the crater edge, which is used to assist in subsequent landing point identification.

[0014] Preferably, when constructing the training set and the validation set based on the public dataset:

[0015] The public datasets used include: SAT-4, SAT-6, LRO, Lunar Orbiter, SELENE / Kaguya, and Lunar DEMs datasets. Select images with similar crater sizes and clear edge information, and adjust the sizes of each image to be consistent to construct a lunar crater detection dataset;

[0016] Divide all the images in the constructed lunar crater detection dataset into a training set and a validation set according to a ratio of 8:2. The training set is used to optimize the model parameters, and the validation set is used to select the optimal model.

[0017] Preferably, each image in the lunar crater detection dataset contains multiple annotation results, where, represents the k-th annotation result, and K is the number of annotations.

[0018] Preferably, the label set {Y (k)} is manually annotated. Each annotation is a 0-1 matrix consistent with the image size, where 1 indicates that the corresponding position in the image is the edge, and 0 indicates that the corresponding position in the image is not the edge.

[0019] Preferably, the UAED detection model includes an encoder, two decoders, two prediction heads, and a Gaussian module. Specifically:

[0020] The image input to the UAED detection model passes through the encoder E and then enters the first path composed of the decoder D1 and the prediction head H1, and the second path composed of the decoder D2 and the prediction head H2 to extract the multi-scale feature matrix of the image. The first path obtains the mean value of the image The second path obtains the variance of the image The mean value and the variance After entering the Gaussian module G, a multivariate Gaussian distribution is constructed And the prediction result is obtained by sampling from this distribution where ε represents a number between 0 and 1 sampled from the standard normal distribution.

[0021] Preferably, the mean value of the outputs of the two paths variance and the prediction result are all matrices with the same size as the input image.

[0022] Preferably, while training the UAED detection model, a loss function is designed according to the model structure, where the loss function includes two parts, expressed as:

[0023]

[0024] where by calculating the variance during training and the label variance σ calculated from the label set to obtain the mean square error; specifically: calculate the variances of K annotations to obtain the annotation image variance σ 2 , and combine the inference of the trained model to obtain the predicted variance 2 Calculate the mean square error of σ and 2 to obtain

[0025] by calculating the binary classification error and mean square error between the prediction result and the label Y sampled from the label set (k) ; specifically: for each image in the training sample, randomly select one from the K annotations as Y (k) , and infer the prediction result according to the trained model Calculate the mean square error of Y (k) and to obtain

[0027] Preferably, when selecting the finally used UAED detection model based on the validation set, the discrimination condition is: select the model with the smallest loss function on the validation set as the optimal model.

[0028] In a second aspect, a terminal device is provided, including:

[0029] A memory for storing instructions executed by at least one processor;

[0030] A processor for executing the instructions stored in the memory to implement the intelligent crater detection method based on the UAED model as described above.

[0031] In a third aspect, a computer-readable storage medium is provided, characterized in that the computer-readable storage medium stores computer instructions, and when the computer instructions run on a computer, the computer is caused to execute the intelligent crater detection method based on the UAED model as described above.

[0032] The present invention has the following advantages compared with the prior art:

[0033] Based on an encoder-decoder structure, the present invention introduces a Gaussian module to consider the differences in different annotations, and adopts a method of ground training combined with on-orbit reasoning to realize on-orbit crater detection in remote sensing images. The detection results are transmitted to the lunar surface lander to assist the lunar surface lander in finding a suitable landing point. On the one hand, compared with traditional crater detection methods, the crater recognition rate and flexibility are improved. On the other hand, a lunar surface crater detection data set is built based on a public data set, solving the problem of missing data samples for lunar surface crater detection. Description of the Drawings

[0034] Figure 1 It is a structural diagram of the UAED edge detection model designed for the present invention. Detailed Embodiments

[0035] With the development of artificial intelligence and neural network technologies, the advantages of methods for lunar surface image processing based on intelligent methods have become prominent. And good results have been achieved in crater detection using intelligent methods.

[0036] The method proposed by the present invention is an intelligent crater detection method based on the UAED (encoder-decoder edge detection integrating uncertainty) model. Compared with traditional crater detection methods, it has the advantages of high recognition rate and high flexibility. It is suitable for on-orbit deployment to assist the lunar surface lander in finding a suitable landing point and improving the landing quality.

[0037] The technical problem solved by the present invention is: to propose an intelligent crater detection method based on the UAED model. Based on the edge detection idea, it adopts the method of ground training and on-orbit reasoning to realize on-orbit crater detection in lunar surface images. The detected crater edge results are transmitted to the lander to help the lander find a more suitable landing location. This method solves the problems of low recognition rate and poor flexibility of traditional crater detection methods. A lunar surface crater detection data set is built based on a public data set, solving the problem of missing data samples for lunar surface crater detection.

[0038] The present invention discloses an intelligent crater detection method based on the UAED model, which is characterized by including the following steps:

[0039] It is generally divided into two stages. In the first stage, the detection model is trained on the ground, and in the second stage, the on-orbit deployment and inference of the detection model are completed on the satellite.

[0040] Step 1: Train the UAED detection model on the ground.

[0041] Step 1.1: Prepare the training and validation datasets for the training model. Based on the publicly available SAT-4, SAT-6, LRO, Lunar Orbiter, SELENE / Kaguya, and Lunar DEMs datasets, select pictures with similar crater sizes and clear edge information to construct a lunar crater detection dataset containing 1100 pictures, and each picture is cropped or expanded to a size of 1024×1024. Divide all images into a training set and a validation set according to a ratio of 8:2. The training dataset is used to optimize the model parameters, and the validation set is used to select the optimal model.

[0042] Step 1.2: Design the structure of the crater detection model. As Figure 1 shown, adopt the edge detection UAED network structure, which includes an encoder, two decoders, two prediction heads, and a Gaussian module.

[0043] Given the input lunar surface image X with a size of [w, h, 3] or [w, h, 1], the encoder E extracts the multi-scale feature matrix of the image and sends the multi-scale feature matrix into two independent decoders D1, D2 and prediction heads H1, H2 to respectively obtain the mean and variance Specifically: the first path composed of the decoder D1 and the prediction head H1, and the second path composed of the decoder D2 and the prediction head H2. The first path obtains the mean of the image, and the second path obtains the variance of the image. The mean and variance are both matrices [w, h, 1] with the same size as the lunar surface image. According to the learned mean and variance construct a multivariate Gaussian distribution where ε represents a number between 0 and 1 sampled from the standard normal distribution, and the prediction result is obtained by sampling from this distribution, which is a matrix [w, h, 1] with the same size as the lunar surface image.

[0044] Step 1.3: Design the loss function according to the model structure. First, obtain the training dataset and the validation dataset through manual annotation for training and selecting the model. For each sample in the dataset, it includes a picture and a label set That is, the K annotation results corresponding to the picture. Each annotation is a 0-1 matrix of the same size as the picture, where 1 indicates that the corresponding position in the picture is an edge, and 0 indicates that the corresponding position in the picture is not an edge.

[0045] During the training process, the loss function is calculated based on the training dataset as and

[0046] By calculating the variance during training and the label variance σ calculated from the label set The mean square error is obtained. The specific calculation method is as follows: for each picture in the training sample, corresponding to K annotations, each annotation is a 0-1 matrix of the same size as the picture, calculate the variance of the K annotations to obtain the annotation variance σ 2 (which is a matrix of the same size as the image), and the predicted variance is obtained by inferring according to the trained model 2 (for the matrix with the same size as the image), calculate the mean square error of σ and 2 to obtain The mean square error of

[0047] By calculating the prediction result and the binary classification error and mean square error of the label Y sampled from the label set (k) The mean square error is obtained. The specific calculation method is as follows: for each picture in the training sample, randomly select one from the K annotations as Y (k) , and the prediction result is obtained by inferring according to the trained model Calculate the mean square deviation of Y (k) and to obtain

[0048] The final optimization objective loss function is and The sum of the two is as follows:

[0049]

[0050] Step 1.4: Train and select the UAED detection model based on the training set and the validation set. Continuously minimize the loss function on the training dataset to optimize the model parameters, and obtain a series of trained UAED detection models. And select the model with the best performance on the validation set as the finally used UAED detection model.

[0051] Among them, the best performance means that the loss function is the smallest on the validation set.

[0052] Step 2: Deploy the UAED detection model on the satellite and on the ground to complete on-orbit inference.

[0053] Step 2.1: Deploy the trained UAED detection model, image preprocessing code, and crater detection inference code on the satellite.

[0054] Step 2.2: Use the UAED detection model in orbit. The satellite generates the original remote sensing images to be processed, inputs them into the image preprocessing and UAED detection inference module codes, calls the UAED detection model file, generates an edge detection matrix, and obtains the crater edge, which is used to assist subsequent landing point recognition.

[0055] In a second aspect, the present invention provides a terminal device, including:

[0056] A memory for storing instructions executed by at least one processor;

[0057] A processor for executing the instructions stored in the memory to implement the intelligent crater detection method based on the UAED model as described above.

[0058] In a third aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, and when the computer instructions are run on a computer, the computer is made to execute the intelligent crater detection method based on the UAED model as described above.

[0059] The content not detailedly described in the specification of the present invention belongs to the prior art well-known to those skilled in the art.

Claims

1. An intelligent crater detection method based on the UAED model, characterized in that include: Construct a lunar crater detection dataset based on public datasets, including training and validation sets; Design the structure of the crater UAED detection model; Design loss function according to model structure; Use the training set to train the UAED detection model. After obtaining a series of trained UAED detection models by minimizing the loss function, select the final UAED detection model based on the validation set. Deploy the trained UAED detection model on board to form a UAED detection inference module; The original remote sensing image is generated on the satellite and input into the UAED detection and reasoning module to detect and obtain the edge of the crater to assist in the subsequent landing site identification.

2. The intelligent crater detection method based on the UAED model according to claim 1, characterized in that: When building training and validation sets based on public datasets: The public datasets used include: SAT-4, SAT-6, LRO, Lunar Orbiter, SELENE / Kaguya and LunarDEMs datasets. Images with similar crater sizes and clear edge information are selected, and the sizes of each image are adjusted to be consistent to construct a lunar crater detection dataset. All images in the constructed lunar crater detection dataset are divided into training set and validation set in a ratio of 8:

2. The training set is used to optimize the model parameters, and the validation set is used to select the optimal model.

3. The intelligent crater detection method based on the UAED model according to claim 1, characterized in that: Each image in the lunar crater detection dataset contains multiple annotation results, among which: Represents the kth annotation result, where K is the number of annotations.

4. The intelligent crater detection method based on the UAED model according to claim 3 is characterized in that: Label set {Y (k) } is manually annotated, and each annotation is a 0-1 matrix consistent with the image size, where 1 indicates that the corresponding position in the image is an edge, and 0 indicates that the corresponding position in the image is not an edge.

5. The intelligent crater detection method based on the UAED model according to claim 1, characterized in that: The UAED detection model consists of an encoder, two decoders, two prediction heads, and a Gaussian module. Specifically: The image input to the UAED detection model passes through the encoder E and enters the multi-scale feature matrix of the extracted image at the same time. The obtained multi-scale feature matrix enters the first path composed of the decoder D1 and the prediction head H1, and the second path composed of the decoder D2 and the prediction head H2. The first path obtains the mean of the image The second pass obtains the variance of the image Mean and variance After entering the Gaussian module G, construct a multivariate Gaussian distribution And get the prediction result by sampling this distribution where ε represents a number between 0 and 1 sampled from a standard normal distribution.

6. The intelligent crater detection method based on the UAED model according to claim 5, characterized in that: The mean of the outputs of the two paths variance And the prediction results The matrices are all the same size as the input image.

7. The intelligent crater detection method based on the UAED model according to claim 3 is characterized in that: While training the UAED detection model, the loss function is designed according to the model structure. It consists of two parts, expressed as: in, By calculating the variance during training With from label set Calculate the label variance σ 2 The mean square error is obtained; specifically: calculate the variance of K annotations and get the variance σ of the annotation image 2 , combined with the trained model inference to get the predicted variance Calculate σ 2 and The mean square error is obtained By calculating the predicted results and the label Y sampled from the label set (k) The binary classification error and mean square error are obtained; Specifically: for each image in the training sample, one is randomly selected from the K annotations as Y (k) , and get the prediction result based on the trained model reasoning Calculate Y (k) and The mean square error is obtained 8. The intelligent crater detection method based on the UAED model according to claim 1, characterized in that: When selecting the final UAED detection model based on the validation set, the judgment condition is: the model with the smallest loss function on the validation set is selected as the optimal model.

9. A terminal device, characterized in that: include: a memory for storing instructions executed by at least one processor; A processor, configured to execute instructions stored in a memory to implement a method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer is enabled to execute the method according to any one of claims 1 to 8.

Images