Sea ice target detection method based on multi-dimensional feature extraction

Abstract

The invention discloses a sea ice target detection method based on multi-dimensional feature extraction, and the method specifically comprises the following steps: S1, data preprocessing: carrying out the denoising of an original sea ice image through a Gaussian filter, carrying out the multi-scale analysis of the sea ice image through an image pyramid, and obtaining sea ice images under different scales; s2, local feature extraction: extracting texture features of each local region by using a gray-level co-occurrence matrix; s3, performing enhanced texture feature processing on the extracted texture features through a directional filter and an enhancement algorithm to obtain final texture feature data; and S4, performing feature fusion on the obtained multi-dimensional features and the texture features to form a comprehensive feature image, and obtaining a classification result of the sea ice target. According to the sea ice target detection method, through multi-scale analysis, image representation is obtained under different scales by using an image pyramid transformation method, changes and expressions of texture features under different scales are captured, and the stability of sea ice target detection is improved.

Description

Technical Field

[0001] The present invention belongs to the field of detection technology, and specifically relates to a sea ice target detection method based on multi-dimensional feature extraction. Background Art

[0002] At present, the most commonly used sea ice target detection method is the full polarization sea ice classification algorithm based on target decomposition features. First, the sea ice data is multi-viewed and filtered to generate a coherence matrix; second, the coherence matrix is ​​decomposed, and the scattering feature parameters are extracted to construct a feature space; third, through statistical correlation analysis, PCA dimensionality reduction is used for highly correlated features to optimize feature combinations; finally, a BP neural network classifier is designed, and the obtained optimized features are used as input to achieve sea ice classification. Its advantage is that it uses full polarization sea ice data to obtain more comprehensive sea ice scattering information, extract features and construct feature space, and optimize feature combinations through statistical correlation and PCA dimensionality reduction, so as to achieve more accurate sea ice classification results. However, there are obvious shortcomings and deficiencies. For example, due to the unstable performance of texture features in different scales and complex sea ice distribution scenes, its role in identifying sea ice is not significant. In addition, this method does not discuss in detail other factors that may affect the accuracy of sea ice classification, such as sea ice shape and structure. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and to provide a sea ice target detection method based on multidimensional feature extraction. The present invention obtains image representations at different scales through multi-scale analysis, using methods such as image pyramid or wavelet transform, captures the changes and manifestations of texture features at different scales, and improves the stability of sea ice target detection.

[0004] In order to achieve the above invention purpose, the technical solution provided by the present invention is as follows:

[0005] A sea ice target detection method based on multi-dimensional feature extraction, the method specifically comprises the following steps:

[0006] S1, data preprocessing, obtaining the original sea ice image, denoising the original sea ice image using a Gaussian filter, performing multi-scale analysis on the denoised sea ice image using an image pyramid, and obtaining sea ice images at different scales;

[0007] S2, local feature extraction, divides the processed sea ice image into multiple local regions, and uses the gray-level co-occurrence matrix to extract texture features for each local region;

[0008] S3, texture feature enhancement and normalization processing, the extracted texture features are enhanced through directional filters and enhancement algorithms, and the enhanced texture features are normalized to obtain the final texture feature data;

[0009] S4, obtaining multi-dimensional features of the sea ice image, and fusion of the obtained multi-dimensional features with the texture features to form a comprehensive feature image, and performing target detection and classification on the comprehensive feature image to obtain the classification result of the sea ice target.

[0010] Furthermore, the Gaussian filter performs denoising on the original sea ice image by: using the Gaussian filter to perform weighted averaging on each pixel in the original sea ice image and surrounding pixels, using the Gaussian kernel function to perform a convolution operation, and calculating the weighted average of each pixel and its surrounding pixels.

[0011] Furthermore, after the original sea ice image is denoised by a Gaussian filter, a Gaussian filter is used to smooth the image of the current scale, and the smoothed sea ice image is downsampled to obtain a sea ice image of the next scale, and the above steps are repeated until a sea ice image of the required scale is obtained.

[0012] Furthermore, the texture feature extraction is specifically as follows: after using the gray level co-occurrence matrix to extract texture features for each local area, the relationship between different gray levels in the sea ice image is measured using an image processing library, and the texture feature vector of each area is obtained by calculating the gray level co-occurrence matrix features of the area.

[0013] Furthermore, the multi-dimensional features of the sea ice image include shape features and color features.

[0014] Furthermore, the comprehensive feature image is subjected to target detection and classification through a support vector machine, and the specific steps are: based on the texture features and multidimensional features extracted from the sea ice image, the features are fused to form a multidimensional feature vector; based on the labeled sea ice target image, the acquired feature vectors are paired with the corresponding target categories to form a training data set; the training data set is divided into a training set and a verification set, and an SVM classifier is used to train the training set, and then the detection accuracy of the trained SVM classifier is verified through the verification set, and finally the target detection result is obtained.

[0015] Based on the above technical solution, the sea ice target detection method based on multi-dimensional feature extraction in the present invention has achieved the following technical advantages through practical application:

[0016] 1. The present invention discloses a sea ice target detection method based on multidimensional feature extraction through multi-scale analysis; the concept of multi-scale analysis is introduced, which means that image representations are obtained at different scales by using methods such as image pyramid or wavelet transform, which can capture the changes and manifestations of texture features at different scales, thereby improving the stability of sea ice target detection.

[0017] 2. The present invention provides a sea ice target detection method based on multidimensional feature extraction. Through local feature extraction, the image is divided into multiple local areas for complex sea ice distribution scenes, and the texture features of each area are extracted. Compared with directly extracting texture features on a global scale, this method can reduce the instability of global texture features and better capture the local texture information of the target.

[0018] 3. The present invention provides a sea ice target detection method based on multi-dimensional feature extraction through texture feature enhancement and normalization; by applying filters and enhancement algorithms to strengthen the texture features and keep them consistent in different sea ice distribution scenarios, which includes using Gaussian filters, median filters, etc. to smooth the image to reduce noise interference and enhance texture features, thereby improving the reliability and stability of texture features.

[0019] 4. The sea ice target detection method based on multidimensional feature extraction of the present invention improves the accuracy and stability of sea ice target detection by combining multidimensional features. This method of comprehensively using multiple features can comprehensively consider information from different aspects, thereby obtaining more reliable and stable target recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of sea ice target detection in a sea ice target detection method based on multi-dimensional feature extraction of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific examples shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0022] like Figure 1 As shown, the present invention belongs to a sea ice target detection method based on multi-dimensional feature extraction, and the method specifically comprises the following steps:

[0023] S1, data preprocessing, obtaining the original sea ice image, denoising the original sea ice image using a Gaussian filter, performing multi-scale analysis on the denoised sea ice image using an image pyramid, and obtaining sea ice images at different scales;

[0024] S2, local feature extraction, divides the processed sea ice image into multiple local regions, and uses the gray-level co-occurrence matrix to extract texture features for each local region;

[0025] S3, texture feature enhancement and normalization processing, the extracted texture features are enhanced through directional filters and enhancement algorithms, and the enhanced texture features are normalized to obtain the final texture feature data;

[0026] S4, obtaining multi-dimensional features of the sea ice image, and fusion of the obtained multi-dimensional features with the texture features to form a comprehensive feature image, and performing target detection and classification on the comprehensive feature image to obtain the classification result of the sea ice target.

[0027] The Gaussian filter performs denoising on the original sea ice image by using a Gaussian filter to perform weighted averaging on each pixel in the original sea ice image and surrounding pixels, and uses a Gaussian kernel function to perform a convolution operation to calculate the weighted average of each pixel and its surrounding pixels.

[0028] After the original sea ice image is denoised by a Gaussian filter, a Gaussian filter is used to smooth the image at the current scale, and the smoothed sea ice image is downsampled to obtain a sea ice image at the next scale. The above steps are repeated until a sea ice image at the required scale is obtained.

[0029] The texture feature extraction is specifically as follows: after extracting texture features from each local area using a gray level co-occurrence matrix, an image processing library is used to measure the relationship between different gray levels in the sea ice image, and the texture feature vector of each area is obtained by calculating the gray level co-occurrence matrix features of the area.

[0030] The multi-dimensional features of the sea ice image include shape features and color features.

[0031] The comprehensive feature image is used for target detection and classification through a support vector machine, and the specific steps are: according to the texture features and multidimensional features extracted from the sea ice image, the features are fused to form a multidimensional feature vector; according to the labeled sea ice target image, the acquired feature vector is paired with the corresponding target category to form a training data set; the training data set is divided into a training set and a verification set, the SVM classifier is used to train through the training set, and the detection accuracy of the trained SVM classifier is verified through the verification set, and finally the target detection result is obtained.

[0032] 1. Through Gaussian filter denoising and multi-scale analysis of image pyramid, the noise interference in the image can be effectively reduced, and sea ice image representations at different scales can be obtained.

[0033] Specific steps:

[0034] (1) Gaussian filter denoising

[0035] 1) Use a Gaussian filter to perform a weighted average of each pixel and its surrounding pixels to reduce high-frequency noise in the image and smooth the image.

[0036] 2) Use the Gaussian kernel function to perform convolution operations and calculate the weighted average of each pixel and its surrounding pixels.

[0037] Key implementation methods:

[0038] A01: Define Gaussian kernel function

[0039] A011: Generate a Gaussian kernel function based on the size and variance of the filter.

[0040] A012: The Gaussian kernel function can be expressed as a two-dimensional matrix, and the value of each element is calculated by the formula of Gaussian distribution. For example, if the filter size is 3x3 and the variance is σ^2, the final generated Gaussian kernel function is: G = (1 / (2*π*σ^2))*exp(-(x^2+y^2) / (2*σ^2))

[0041] A02: Perform filtering operation

[0042] A021: For each pixel, the size of the neighborhood area around it is taken as the size of the filter.

[0043] A022: Multiply the Gaussian kernel function by the pixel values ​​in the neighborhood element by element and sum them to get the weighted result.

[0044] A023: Assign the weighted result to the original pixel location, and the smoothing operation is completed.

[0045] A024: Repeat the above steps and perform filtering operations on all pixels.

[0046] (2) Multi-scale analysis of image pyramid

[0047] Based on the Gaussian pyramid of Gaussian filter, images of different scales are obtained by smoothing and downsampling the image in sequence. The following steps are performed in a loop:

[0048] 1) Use a Gaussian filter to smooth the image at the current scale.

[0049] 2) Downsample the smoothed image to obtain an image of the next scale.

[0050] 3) Take the image of the next scale as input and continue the above steps until the desired scale level is reached.

[0051] 2. Use grid division to divide the image into multiple local areas and extract the texture features of each area. This can better capture the local texture information of sea ice targets and help accurate target detection and classification;

[0052] Specific steps:

[0053] (1) Grid division

[0054] 1) Use loops and index operations in a programming language to divide the image into multiple local regions.

[0055] 2) The size and shape of the grid are defined according to actual needs, such as square or rectangle, and the size and spacing are adjusted as needed.

[0056] (2) Regional texture feature extraction

[0057] 1) Use the gray-level co-occurrence matrix (GLCM) to extract texture features for each local area.

[0058] 2) Use image processing libraries (such as OpenCV) to measure the relationship between different gray levels in the image, including contrast, energy, correlation, etc. By calculating the GLCM features of each region, the texture feature vector of the region is obtained.

[0059] Key implementation methods:

[0060] A01: Algorithm that uses loops and index operations to divide an image into multiple local regions. First, load the original image. Then use the partition_image function to divide the image into multiple local regions, where the grid_size parameter defines the size of each region. Through loops and index operations, the image is divided into square local regions and stored in the regions list. Finally, use a loop to traverse and display each local region.

[0061] A02: Algorithm for extracting texture features using gray-level co-occurrence matrix (GLCM). First, load the original image. Then use the partition_image function to divide the image into multiple local regions and get a list of regions. Next, use the extract_texture_features function to extract texture features for each local region. By calculating the gray-level co-occurrence matrix (GLCM) and using the compareHist function in OpenCV, we can get the values ​​of different texture features, such as contrast, energy, and correlation. Finally, store the texture feature vector in the texture_features list and print it out.

[0062] 3. Introducing directional filters and enhancement algorithms to process local features can enhance the recognition of texture features, and make the features comparable and consistent through normalization;

[0063] The specific steps and key implementation methods are as follows:

[0064] (1) Introducing directional filters

[0065] The local features are processed using Gabor filters and the functions provided by the corresponding image processing library (such as OpenCV) are called to achieve this.

[0066] A01: Load the original image.

[0067] A02: Definition: The parameters of the Gabor filter that needs to be used, such as kernel size (kernel_size), direction (theta), standard deviation (sigma), wavelength (lambd) and gamma factor (gamma).

[0068] A03: Create a Gabor filter by calling the getGaborKernel function in the OpenCV library.

[0069] A04: Use the filter2D function to apply the Gabor filter to the original image and get the filtered result. Use the imshow function to display the original image and the filtered result.

[0070] (2) Enhanced algorithm

[0071] The histogram equalization enhancement algorithm is used to enhance texture features.

[0072] A01: Load the original image.

[0073] A02: Use the enhance_texture_features function to apply the histogram equalization enhancement algorithm to the image, which can improve the contrast of the image and enhance the texture features.

[0074] A03: Perform histogram equalization on the image by calling the equalizeHist function in the OpenCV library.

[0075] A04: Use the imshow function to display the original image and the enhanced image.

[0076] (3) Normalization

[0077] The filtered features are normalized through mathematical library functions of programming languages, such as linear scaling, standardization, maximum and minimum value normalization, etc., to ensure that the feature value range between different samples is unified and the features are comparable and consistent.

[0078] A01: Define three feature normalization functions: linear scaling (linear_scaling), standardization (standardization) and maximum and minimum normalization (min_max_normalization);

[0079] A02: Given an original feature array, normalize the feature array by calling the corresponding normalization function and obtain the normalized feature array;

[0080] A03: Print and display the normalized feature array.

[0081] 4. By combining multiple features such as texture, shape, and color, different aspects of sea ice targets can be comprehensively considered, thereby improving the accuracy and stability of target detection and classification.

[0082] The specific steps and key implementation methods are as follows:

[0083] (1) Feature extraction

[0084] 1) Texture features: Call the image processing library (such as OpenCV) and use the GLCM method to extract the texture features of the image.

[0085] A01: Load the original image;

[0086] A02: Use cv2.calcGLCM function to calculate the gray-level co-occurrence matrix (GLCM). The parameters of this function include the image for which GLCM is to be calculated, the channel index list (just use [0] in grayscale images), and the number of grayscale levels (for example, 256 levels);

[0087] A03: Extract texture features from GLCM using cv2.compareHist function by calculating histogram comparison values ​​(i.e. contrast, energy and correlation) between different texture features;

[0088] A04: Print texture features.

[0089] 2) Shape features: Use contour descriptors such as Hu moments and bounding boxes to extract the shape features of the image.

[0090] A01: Load the original image;

[0091] A02: Convert the image into a grayscale image and perform binarization processing on it;

[0092] A03: Use cv2.findContours function to extract contours. In this example, we only selected the first contour found. You can select a specific contour at your own discretion.

[0093] A04: Use cv2.moments to calculate the moments of the contour. By passing these moments to the cv2.HuMoments function, the Hu invariant moments can be calculated;

[0094] A05: Print the Hu moment features and the calculated bounding box information.

[0095] 3) Color features: Use color histogram to extract the color features of the image.

[0096] A01: Load the original image;

[0097] A02: Convert the image from BGR color space to RGB color space, because when displaying images in Matplotlib, the RGB color channel order is used by default;

[0098] A03: Use the cv2.calcHist function to calculate the color histogram of the image. Select all three color channels (R, G, B) and specify the size and range of the color histogram. The normalized color histogram can be obtained by normalizing and flattening the histogram into a one-dimensional vector;

[0099] A04: Use the Matplotlib library to draw a normalized color histogram.

[0100] (2) Feature Fusion

[0101] The different extracted features are spliced ​​or combined to form a comprehensive feature image, which is then input into the classifier for target detection and classification.

[0102] A01: Load image;

[0103] A02: Extract texture features, shape features and color features respectively;

[0104] A03: Concatenate or combine these features as needed to form a comprehensive feature image. Usually, the concatenation operation can be completed by using the concatenate function in the numpy library;

[0105] A04: Load the trained classifier model and use the classifier to perform target detection and classification on the comprehensive feature image.

[0106] 5. Using support vector machine (SVM) as a classifier and utilizing the extracted multi-dimensional features to detect and classify sea ice targets, accurate target segmentation and classification results can be obtained.

[0107] Specific steps:

[0108] (1) Feature vector preparation

[0109] 1) According to the feature extraction steps mentioned above, multi-dimensional features such as texture, shape, color, etc. are extracted from the sea ice image;

[0110] 2) Combine the feature extraction results of each target into a multi-dimensional feature vector.

[0111] (2) Dataset preparation

[0112] According to the labeled sea ice target images, the extracted feature vectors are paired with their corresponding target categories (such as sea ice or non-sea ice) to form a training dataset.

[0113] (3) Data preprocessing

[0114] 1) Preprocess the training data set, such as feature normalization and data partitioning.

[0115] 2) Feature normalization ensures comparability between different features.

[0116] 3) The data is divided into a training set and a validation set for training the classifier and evaluating the classification accuracy.

[0117] (4) SVM classifier training

[0118] 1) Use SVM as a classifier for training;

[0119] 2) In the training phase, the model is trained using the feature vectors of the training data set and the corresponding category labels;

[0120] 3) SVM is a supervised learning algorithm that learns a decision boundary based on the differences between feature vectors and category labels for target classification.

[0121] (5) Object detection and classification:

[0122] 1) Extract features from the sea ice image data to be detected and construct a test data set;

[0123] 2) Use the trained SVM model to classify the targets in the test data set;

[0124] 3) For unknown samples, the corresponding target segmentation and classification results can be obtained based on the classification results.

[0125] Key implementation methods:

[0126] A01: Load sea ice images and corresponding labels. Load images by looping through them and extract corresponding features.

[0127] A02: Use the train_test_split function to divide the data set into training set and test set;

[0128] A03: Extract features of the training set and test set, and initialize the SVM classifier. Use the fit function to train the training set;

[0129] A04: Make predictions on the test set and calculate performance indicators such as accuracy, precision, recall, and F1-Score.

[0130] The key point lies in the innovative feature extraction method and the combination of multi-dimensional features. By introducing technologies such as multi-scale analysis, local feature extraction and texture feature enhancement normalization, this method can more effectively capture the texture information of sea ice targets and combine other features for comprehensive analysis, thereby improving the accuracy and stability of detection.

[0131] Example

[0132] Background: In the Arctic, changes in sea ice are of great significance to climate research. For example, glacier melting and sea ice coverage are related to global warming. In order to monitor the distribution and evolution of sea ice, researchers need an accurate and stable sea ice target detection method.

[0133] The following is a practical scenario of sea ice target detection method based on multidimensional feature extraction in the Arctic region:

[0134] 1. Multi-scale analysis: Researchers used drones to capture images of sea ice in the Arctic. They used high-resolution cameras equipped with drones to capture a series of images at different flight altitudes. Then, for each image, they constructed an image pyramid, gradually reducing the resolution of the image to obtain image representations at different scales.

[0135] 2. Local feature extraction: In order to capture the complex sea ice distribution scenes, researchers divide each image into small local areas. They use image segmentation algorithms to divide the image into many grid cells and extract the texture features of each cell. This can reduce the instability of texture features across the entire image and better capture the local texture information of the sea ice target.

[0136] 3. Texture feature enhancement and normalization: To strengthen texture features and make them more consistent, researchers applied filters and enhancement algorithms to each local area. They used Gaussian filters to smooth images, remove noise interference, and enhance texture features. In addition, they used adaptive histogram equalization technology to adjust the contrast of the image to make the texture features and targets appear clearer.

[0137] 4. Combining other features: In addition to texture features, researchers also combined shape features and color features to improve the accuracy and stability of sea ice target detection. They extracted contour descriptors based on the shape features of each target, and extracted color histograms based on the color features. By fusing these multi-dimensional features, they comprehensively considered the target's texture, shape, color and other information to obtain more accurate and stable sea ice target recognition results.

[0138] In this actual scenario, the sea ice target detection method based on multidimensional feature extraction enables researchers to comprehensively analyze sea ice information in the Arctic. By applying multi-scale analysis, local feature extraction, texture feature enhancement normalization and combining other features, they can more accurately monitor the distribution and changes of sea ice, providing strong support for climate research.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. A sea ice target detection method based on multidimensional feature extraction, characterized in that: The method specifically comprises the following steps: S1, data preprocessing, obtaining the original sea ice image, denoising the original sea ice image using a Gaussian filter, performing multi-scale analysis on the denoised sea ice image using an image pyramid, and obtaining sea ice images at different scales; S2, local feature extraction, divides the processed sea ice image into multiple local regions, and uses the gray-level co-occurrence matrix to extract texture features for each local region; S3, texture feature enhancement and normalization processing, the extracted texture features are enhanced through directional filters and enhancement algorithms, and the enhanced texture features are normalized to obtain the final texture feature data; S4, obtaining multi-dimensional features of the sea ice image, and fusion of the obtained multi-dimensional features with the texture features to form a comprehensive feature image, and performing target detection and classification on the comprehensive feature image to obtain the classification result of the sea ice target.

2. The sea ice target detection method based on multidimensional feature extraction according to claim 1 is characterized in that: The Gaussian filter performs denoising on the original sea ice image by using a Gaussian filter to perform weighted averaging on each pixel in the original sea ice image and surrounding pixels, and uses a Gaussian kernel function to perform a convolution operation to calculate the weighted average of each pixel and its surrounding pixels.

3. The sea ice target detection method based on multidimensional feature extraction according to claim 2 is characterized in that: After the original sea ice image is denoised by a Gaussian filter, a Gaussian filter is used to smooth the image at the current scale, and the smoothed sea ice image is downsampled to obtain a sea ice image at the next scale. The above steps are repeated until a sea ice image at the required scale is obtained.

4. The sea ice target detection method based on multidimensional feature extraction according to claim 1 is characterized in that: The texture feature extraction is specifically as follows: after extracting texture features from each local area using a gray level co-occurrence matrix, an image processing library is used to measure the relationship between different gray levels in the sea ice image, and the texture feature vector of each area is obtained by calculating the gray level co-occurrence matrix features of the area.

5. The method for sea ice target detection based on multidimensional feature extraction according to claim 1, characterized in that: The multi-dimensional features of the sea ice image include shape features and color features.

6. The method for sea ice target detection based on multidimensional feature extraction according to claim 1, characterized in that: The comprehensive feature image is used for target detection and classification by a support vector machine, and the specific steps are: according to the texture features and multi-dimensional features extracted from the sea ice image, the features are fused to form a multi-dimensional feature vector; according to the labeled sea ice target image, the acquired feature vector is paired with the corresponding target category to form a training data set; The training data set is divided into a training set and a validation set. The SVM classifier is trained with the training set, and the detection accuracy of the trained SVM classifier is verified with the validation set to finally obtain the target detection result.

Images