A garbage remote sensing identification method and system based on high-order singular value decomposition and distributed processing

Abstract

The present invention discloses a method and system for remote sensing identification of solid waste based on high-order singular value decomposition and distributed processing, including Step 1: receiving remote sensing image data of solid waste, and representing the image data with tensors. Step 2: distributed processing, forward slicing the obtained tensors into matrices through distribution and grouping them, and recombining the matrices after decomposition calculation into approximate tensors to reduce the time cost during noise reduction. Step 3: denoising the obtained approximate tensors using the adaptive residual method. Step 4: feature fusion and target recognition: analyzing the extracted features, and studying the influence degree of the features of different regions in the remote sensing image of solid waste on the classification result according to the gradient intensity and position information of the feature maps, so as to improve the recognition accuracy. The present invention can quickly process a large amount of image data and can accurately identify solid waste, and the obtained feature information is more accurate, achieving the technical purpose of more accurately identifying and classifying solid waste in remote sensing images.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing target recognition based on machine learning, and specifically relates to a method and system for remote sensing recognition of garbage based on high-order singular value decomposition and distributed processing. Background Art

[0002] In recent years, with the acceleration of the urbanization process, the problem of solid waste has become increasingly prominent. Solid waste refers to any domestic, industrial, and construction waste that can no longer be used, and is usually classified into recyclables, hazardous waste, and other waste. Solid waste has a serious impact on the environment and human health. Among them, hazardous waste contains various toxic and harmful substances, such as heavy metals, hazardous chemicals, medical waste, etc. If not properly treated, it will lead to problems such as groundwater pollution, land degradation, and air pollution, thus endangering human health and the ecological environment.

[0003] Remote sensing images of solid waste usually have low resolution, especially when analyzing large geographical areas. This limits the visibility of details and features in the images and increases the difficulty of target recognition. Remote sensing images of solid waste may be affected by various noises and distortions, such as cloud cover, atmospheric turbulence, and the noise of the sensor itself. These factors will affect the quality and readability of the images and reduce the accuracy of target recognition. Remote sensing images of solid waste are usually taken from different perspectives and heights, resulting in great variations in the scale and appearance of objects in the images. This requires target recognition algorithms to be robust to scale and perspective changes. The target objects in remote sensing images of solid waste are often blocked by other objects or landforms, and there may be deformations in the images. These blockages and deformations will make the targets in the images blurred or incomplete, thus increasing the difficulty of target recognition. Therefore, the main difficulties of traditional remote sensing images are summarized in the following aspects:

[0004] (1) Remote sensing images have characteristics such as high dimension, multi-scale, and multi-spectral features. Therefore, it is necessary to consider how to extract effective feature information during the processing.

[0005] (2) The amount of data in remote sensing images is large, including differences in acquisition time, spatial resolution, number of bands, etc. Therefore, it is necessary to consider how to perform effective data compression and processing to reduce the time cost.

[0006] (3) Remote sensing target recognition needs to consider the influence of various factors such as ground object coverage type, terrain, vegetation, soil, etc. Therefore, it is necessary to fully consider the influence of these factors on the recognition results.

[0007] (4) Remote sensing target recognition tasks require a large amount of labeled data to train models. However, due to reasons such as high image resolution, large quantity, and multiple categories, the acquisition cost of labeled data is relatively high. Therefore, it is necessary to explore effective semi-supervised or unsupervised learning methods.

[0008] Machine learning methods and deep learning methods have been widely applied in the field of remote sensing image processing of solid waste, and good results have been achieved in remote sensing target recognition. However, there are still the following technical problems:

[0009] First, traditional machine learning will show less than ideal results when dealing with high-dimensional and high-precision remote sensing images.

[0010] Second, in the recognition of solid waste, errors will occur when there are multiple targets in the image for simultaneous recognition or when there are too many similar interference items.

[0011] Third, the high data dimension and clarity bring problems such as high computational complexity and easy occurrence of computational explosion when processing large batches of data. Summary of the Invention

[0012] Aiming at the problems of large computational loss and unsatisfactory accuracy in the remote sensing target recognition of solid waste with high dimensions and multiple targets, the present invention proposes a garbage remote sensing recognition method and system based on high-order singular value decomposition and distributed processing. This method optimizes the computational explosion problem during tensor compression and improves the processing speed by performing forward slicing of the tensor and optimally allocating it to each computing node. In the recognition, a feature fusion algorithm is introduced to fuse features at different stages for target recognition, improving the recognition accuracy. At the same time, this algorithm is a weakly supervised algorithm, which optimizes the situation where traditional machine learning must be processed with labels to a certain extent.

[0013] To achieve the above technical objectives, the present invention adopts the following specific technical solutions:

[0014] A garbage remote sensing recognition method based on high-order singular value decomposition and distributed processing, comprising the following steps:

[0015] Step 1: Acquisition of remote sensing image data

[0016] Receive solid waste remote sensing image data and represent the image data with a tensor;

[0017] Step 2: Distributed processing

[0018] Through distribution, forward slice the tensor obtained in Step 1 into matrices and group them, and then recombine the matrices after decomposition calculation into an approximate tensor to reduce the time cost during noise reduction and avoid the computational explosion problem;

[0019] Step 3: Image denoising

[0020] Denoise the approximate tensor obtained in step 2 using the adaptive residual method, establish an adaptive residual function and combine multiple iterations to improve the denoising effect;

[0021] Step 4: Feature fusion and target recognition

[0022] Segment the image after noise reduction in step 3 according to the patch size, divide it into three stages according to the image size, segment it into multiple patch blocks, extract features from each patch block, start calculating from the stage with the largest number of patch blocks, then transfer the feature results calculated in the current stage to the next stage for feature fusion, and finally perform target recognition after feature fusion in each stage to obtain the matching area of solid waste in the satellite remote sensing image.

[0023] In step 1, the obtained satellite image data is converted into a tensor using pytorch.

[0024] Step 2 specifically includes:

[0025] 2.1) According to the tensor in step 1, perform an optimal division on the satellite solid waste remote sensing image dataset along the time series, count the number of non-zero elements in the forward matrix of each time series. Assume that the total number of non-zero elements is n, and assume that there are k processing nodes at the same time. Then the division index is n / k, that is, the tensor composed of the first m matrices whose sum of non-zero elements reaches the index is assigned to the first node and so on;

[0026] 2.2) Each node performs a singular value decomposition on the processed sub-tensor according to the forward slice matrix, such as the matrix After decomposition, three decomposed matrices are obtained , where is an orthogonal matrix, is a diagonal matrix, and the elements on the diagonal are singular values, that is, the eigenvalues of the matrix . According to the number of singular values to be retained , intercept the first diagonal elements of the diagonal matrix to calculate the new approximate matrix , and then combine the obtained new approximate matrix into the approximate tensor of the current node, and finally re-represent it as an image dataset;

[0027] The distributed processing of data reduces the computational time cost and avoids the problem of computational explosion that may occur in the denoising step.

[0028] Step 3 is specifically:

[0029] By defining an adaptive residual denoising function and performing multiple iterative calculations to improve the denoising effect; the residual image is the difference between the original noisy image and the denoised image. Assume the original noisy image is and the denoised image is , and the residual image is . To improve the denoising effect, the image after noise reduction processing is added with the product of the residual image and the ratio of noise reduction processing. Define the ratio of noise reduction processing ; Assume the image result obtained after k times of denoising is , is the residual image obtained after k times of denoising; The processing formula for each step is ; The final denoising result is obtained through 15 times of iterative processing.

[0030] Step 4 specifically includes the following sub-steps:

[0031] 4.1) According to the image size, at ratios of 1 / 16, 1 / 4, and 1 / 1, divide the image into three stages in patch blocks, and use a loop to traverse the image and divide it into blocks;

[0032] 4.2) Define a convolutional neural network model. This convolutional neural network model includes a global average pooling layer. The global average pooling layer converts the output feature map of the last convolutional layer into a feature vector. Use a labeled training data set to train the convolutional neural network model and adjust its weights;

[0033] 4.3) Use the trained convolutional neural network model to calculate the two-dimensional heat map of each patch block. Among them, each pixel represents the contribution degree of the pixel to a specific category. The generation process of the two-dimensional heat map CAM is realized through the backpropagation method. Start calculating from the stage with the largest number of patch blocks, then pass the feature results calculated in the current stage to the next stage for feature fusion, and finally perform target recognition after the feature fusion of each stage is completed to obtain the solid waste garbage matching area in the satellite remote sensing image;

[0034] Through the means of feature fusion, good accuracy can be achieved in the face of a large amount of data interference information and multiple recognition targets.

[0035] The present invention further discloses a garbage remote sensing recognition system, which uses the garbage remote sensing recognition method based on high-order singular value decomposition and distributed processing for recognition.

[0036] The beneficial effects of the present invention are:

[0037] A garbage remote sensing identification method based on high-order singular value decomposition and distributed processing. This method optimizes the computational explosion problem during tensor compression and improves the processing speed by performing forward slicing of the tensor and optimally allocating it to each computing node. In the identification, a feature fusion algorithm is introduced to fuse features at different stages for target identification, improving the identification accuracy. At the same time, this algorithm is a weakly supervised algorithm, which optimizes the situation where traditional machine learning must be processed with labels to a certain extent.

[0038] High efficiency: The traditional matrix calculation for image data has a high time complexity. In the present invention, the incoming image data is converted into a high-order tensor, and then the tensor is transformed and classified according to the rank of the slice, and the processing tasks of the image are assigned to different task nodes to improve the processing rate.

[0039] Robustness: Since the remote sensing image data of solid waste may be affected by many factors such as bad weather, surrounding buildings, etc., traditional machine learning will be greatly disturbed when considering these factors. Therefore, in the present invention, feature extraction is performed based on global average pooling during feature extraction, suppressing the detailed information in the image and retaining the main global features. The influence of noise is reduced to a certain extent through the augmentation and enhancement of training data, and the weights of occluded and unobvious regions are reduced. Overall, the robustness of the algorithm is improved to cope with special situations in various cases.

[0040] Accuracy: Different from feature extraction under traditional machine learning, in the present invention, the image data is divided into three times to extract feature values, and then feature fusion is performed, improving the accuracy of the extracted feature values. Brief Description of the Drawings

[0041] Figure 1 is the overall architecture of the garbage remote sensing identification method based on high-order singular value decomposition and distributed processing of the present invention;

[0042] Figure 2 is the flowchart of feature extraction in the garbage remote sensing identification method based on high-order singular value decomposition and distributed processing of the present invention.

[0043] Figure 3 is the comparison chart of the image accuracy of the present invention compared with the traditional denoising method;

[0044] Figures 4 to 6 is the performance result of the present invention through feature fusion in the situation of multiple targets and many interferences in the image. Detailed Embodiment

[0045] I. System Structure

[0046] The architecture of this method can be divided into four parts: the data processing part, the distributed processing part, the data denoising part, and the feature extraction and target recognition part. Its overall architecture is as Figure 1 shown.

[0047] Data processing part: First, convert the image data into a tensor model, which helps to represent data features in different dimensions and is more convenient for storing a large amount of data. And label the dataset for easy training.

[0048] Distributed processing part: The obtained approximate tensor is divided into multiple slice matrices according to the forward slices of the approximate tensor. According to the number of non-zero elements in the slice matrix, the slice matrices are grouped and distributed to each node for distributed processing. Each node performs singular value decomposition on the assigned matrix and finally recombines the decomposed matrices into a new approximate tensor to obtain the compressed image.

[0049] Data denoising part: And use the adaptive residual method to denoise the new approximate tensor for the image.

[0050] Feature extraction and target recognition part: Use the preprocessed dataset to train the selected model. During the training process, optimize the model by minimizing the loss function so that it can accurately predict image classification. Obtain the feature maps related to the classification of solid waste remote sensing images. These feature maps can be used as the feature values of the solid waste remote sensing images, reflecting the contribution degree of different regions in the image to the classification result.

[0051] When processing images, we divide the images into three different ways: 1 / 16, 1 / 4, and 1 / 1 for feature extraction and perform feature fusion in the next step. By analyzing the extracted features, we can study the influence degree of the features of different regions in the solid waste remote sensing image on the classification result according to information such as the gradient intensity and position of the feature maps.

[0052] II. Method flow

[0053] 1. Acquisition of remote sensing image data

[0054] Data processing part: Collect the images taken of solid waste, convert them into tensor data, and label the dataset.

[0055] Distributed processing part: The obtained tensor is divided into multiple slice matrices according to the forward slices of the approximate tensor. According to the number of non-zero elements in the slice matrix, the slice matrices are grouped and distributed to each node for distributed processing. Assume the number of non-zero elements is n, and we set the current total number of processing nodes to k. Then the number of non-zero elements in the matrix processed by one node is approximately n / k. Perform singular value decomposition on each slice matrix and retain the first singular values, The number accounts for 10% - 20% of the total number of singular values. Finally, the decomposed matrix is recombined into a new approximate tensor to obtain the compressed image, avoiding the situation during noise reduction processing.

[0056] Data noise reduction part: By defining an adaptive residual denoising function and performing multiple iterative calculations to improve the denoising effect.

[0057] The residual image is the difference between the original noisy image and the denoised image. Assume the original image is , and the denoised image is , then the residual image at this time is . To improve the denoising effect, we multiply the image obtained through noise reduction processing by the product of the ratio of the residual image to the noise reduction processing. We define the ratio of noise reduction processing as . Assume the image result obtained after k times of denoising is , is the residual image obtained after k times of denoising. The processing formula for each step is .

[0058] The final denoising result is obtained through 15 times of iterative processing.

[0059] After comparison, it is found that it has a better performance in image accuracy compared to traditional denoising methods, as shown in Figure 3 .

[0060] Feature extraction and target recognition part: Use the preprocessed dataset to train the selected model. During the training process, the model is optimized by minimizing the loss function so that it can accurately predict image classification. Obtain the feature maps related to the remote sensing image classification of solid waste. These feature maps can be used as the feature values of the remote sensing image of solid waste, reflecting the contribution degree of different regions in the image to the classification result.

[0061] During image processing, we divide the denoised result image obtained in the previous step into three different ways: 1 / 16, 1 / 4, and 1 / 1 for feature extraction and perform feature fusion in the next step. By analyzing the extracted features, we can study the influence degree of the features of different regions in the remote sensing image of solid waste on the classification result according to information such as the gradient intensity and position of the feature maps. Through the method of feature fusion, accurate results are obtained in the case of multiple targets and many interferences in the image, as shown in Figures 4 to 6 .

Claims

1. A garbage remote sensing recognition method based on high-order singular value decomposition and distributed processing, characterized in that, It includes the following steps: Step 1: Acquisition of remote sensing image data Receive the remote sensing image data of solid waste, and represent the image data with tensors; Step 2: Distributed processing Forward-slice the tensor obtained in Step 1 into matrices through distribution and then group them, and decompose and recombine the matrices to form an approximate tensor to reduce the time cost during noise reduction and avoid the problem of computational explosion; Step 3: Image denoising Denoise the approximate tensor obtained in Step 2 using the adaptive residual method, and establish an adaptive residual function and combine multiple iterations to improve the denoising effect; Step 4: Feature fusion and target recognition Segment the image after noise reduction in Step 3 according to the patch size, divide it into three stages according to the image size, segment it into multiple patch blocks, extract features from each patch block, start calculating from the stage with the largest number of patch blocks, then pass the feature results calculated in the current stage to the next stage for feature fusion, and finally perform target recognition after feature fusion in each stage to obtain the matching area of solid waste in the satellite remote sensing image; Specifically, Step 3 is as follows: By defining an adaptive residual denoising function and performing multiple iterative calculations to improve the denoising effect; the residual image is the difference between the original noisy image and the denoised image. Assume the original noisy image is , the denoised image is , and the residual image is . The image after noise reduction is added with the product of the residual image and the ratio of noise reduction processing. Define the ratio of noise reduction processing as . Assume the image result obtained after k times of denoising is , is the residual image obtained after k times of denoising; The processing formula for each step is obtained as ; The final denoising result is obtained through 15 iterations of processing.

2. The method for remotely sensing and identifying garbage based on high-order singular value decomposition and distributed processing according to claim 1, wherein In Step 1, convert the obtained satellite image data into a tensor using pytorch.

3. The garbage remote sensing recognition method based on high-order singular value decomposition and distributed processing according to claim 1, characterized in that Specifically, Step 2 includes: 2.1) According to the tensor in Step 1, perform an optimal division of the satellite solid waste remote sensing image dataset along the time series, count the number of non-zero elements in the forward matrix of each time series. Assume the total number of non-zero elements is n, and assume there are k processing nodes at the same time. Then the division index is n / k, that is, the tensor composed of the first m matrices whose total number of non-zero elements reaches the index is assigned to the first node and so on; 2.2) Each node performs singular value decomposition on the processed sub-tensor according to the forward slicing matrix. The matrix After decomposition, three decomposed matrices are obtained , where is an orthogonal matrix, is a diagonal matrix, and the elements on the diagonal are singular values, that is, the eigenvalues of the matrix . According to the number of singular values to be retained , the first diagonal elements of the diagonal matrix are intercepted to calculate a new approximate matrix . Then, the obtained new approximate matrix is combined into the approximate tensor of the current node, and finally, it is re-represented as an image data set.

4. The method for remotely sensing and identifying garbage based on high-order singular value decomposition and distributed processing according to claim 1, wherein Specifically, Step 4 includes the following sub-steps: 4.1) According to the image size, divide the image into three stages according to the ratios of 1 / 16, 1 / 4, and 1 / 1 in terms of patch blocks, and use a loop to traverse the image and divide it into blocks; 4.2) Define a convolutional neural network model. This convolutional neural network model includes a global average pooling layer. The global average pooling layer converts the output feature map of the last convolutional layer into a feature vector, and use the labeled training dataset to train the convolutional neural network model and adjust its weights; 4.3) Use the trained convolutional neural network model to calculate the two-dimensional heat map of each patch block. Among them, each pixel represents the contribution degree of the pixel to a specific category. The generation process of the two-dimensional heat map CAM is realized through the backpropagation method. Start calculating from the stage with the largest number of patch blocks, then pass the feature results calculated in the current stage to the next stage for feature fusion, and finally perform target recognition after feature fusion in each stage to obtain the matching area of solid waste in the satellite remote sensing image.

5. A garbage remote sensing identification system, characterized in that, Perform recognition using the garbage remote sensing recognition method based on high-order singular value decomposition and distributed processing as described in any one of claims 1 to 4.

Images