CN-115984697-B - SAR image oil spill detection method

CN115984697BCN 115984697 BCN115984697 BCN 115984697BCN-115984697-B

Abstract

The invention relates to the field of synthetic aperture radar image oil spill segmentation, in particular to an SAR image oil spill detection method, which adopts a data enhancement technology to amplify a data set, fully trains parameters in a model, improves generalization capability and robustness of the model, reduces overfitting, uses Wavelet Threshold Transformation (WTT) to eliminate noise and highlight global features, adds a residual module (ResidualModel), a SE attention module and a feature pyramid module (ASPP) with cavity convolution in an original U-Net network, effectively extracts useful target features existing in an SAR image, reduces the number of parameters in the model, shortens training time, optimizes model performance, and embeds a coordination attention module (CA) at a jump joint to eliminate redundant low-layer features. Compared with the prior art, the method provided by the invention has better effect.

Inventors

WEI XUEYUN
CHEN SIYUAN
ZHANG ZHENKAI
ZHENG WEI
JIN BIAO
XI CAIPING
SHANG SHANG

Assignees

江苏科技大学

Dates

Publication Date: 20260512
Application Date: 20230104

Claims (6)

1. The SAR image oil spill detection method is characterized by comprising the following steps of: Step 1, selecting pictures in an existing oil spill detection database as a training set; step 2, preprocessing the training set selected in the step 1, namely randomly cutting and splicing the training set; step 3, denoising the training set obtained in the step 2 by wavelet threshold transformation; Step 4, building a feature fusion U-Net model based on a coordinated attention mechanism, randomly selecting 50% training set pictures obtained in the step 2 and the step 3 as the input of a network model, and carrying out global feature extraction and fusion; step 5, using a residual error module, and adding an extrusion and excitation module after part of the residual error module to enable the model to autonomously learn the weight coefficient of each channel; step 6, embedding a coordination attention mechanism module at the jump joint to eliminate redundant information; step 7, a space pyramid structure with cavity convolution is used as a bottom network, and the receptive field is increased and meanwhile, wider characteristic information is extracted; Step 8, inputting the selected test set into a trained feature fusion U-Net model based on a coordinated attention mechanism for testing, and obtaining a network segmentation result; The specific flow of the step 1 is that 750 pictures are selected from the European space agency oil spill detection dataset to serve as training sets, 250 pictures are selected as test sets, 90% of the training sets are randomly selected to be used for experimental training, and 10% of the training sets are used for experimental verification; The wavelet threshold function proposed in the step 3 introduces variables so that the whole function is continuously conductive and relatively smooth at the threshold, and the wavelet threshold function is as follows: (1) wherein m and k are variables, m.epsilon.0, 1, and k is a positive integer.
2. The SAR image oil spill detection method according to claim 1, wherein the specific flow of the step 2 is that the training set selected in the step 1 is preprocessed, namely, the training set is randomly cut and spliced, wherein the original SAR image format is 1250×650, the preprocessed data set is remolded into 256×256 images, the number of characteristic channels is 3, the specific data enhancement operation comprises the steps that the size of the randomly cut images is 0.2-0.4 times of the original image size, the cutting length-width ratio is 1:2, and the cut images are spliced two by two and are remolded into 256×256×3 format.
3. The SAR image oil spill detection method according to claim 2, wherein the specific flow of step 4 is as follows: 4.1, fusing U-Net model based on coordination attention mechanism features, adopting 17 convolution layers, 4 downsampling layers, 4 upsampling layers and 4 clipping and copying layers, wherein the format of an input image is 256 multiplied by 3, the convolution layers are two by two to form a residual block, each convolution layer adopts a convolution kernel with the size of 3 multiplied by 3, an activation function is swish, all zero filling is adopted, an extrusion excitation module is placed after the residual block to improve the distinguishing power of the model to the features, the downsampling layer adopts a maximum pooling method, the window size of sampling is 2 multiplied by 2, the step size is 2, a pyramid structure with cavity convolution is placed at the bottom layer of the network, the receptive field is increased, the upsampling layer adopts inverse convolution, the convolution kernel size is 3 multiplied by 3, the step size is 2, all zero filling is used, the coordination attention module is embedded into jump connection, the redundant features are eliminated, and finally, the convolution layer with the size of 1 multiplied by 1 is reduced to the dimension, and the segmentation result is output; And 4.2, randomly selecting 50% training set images obtained in the steps 2 and 3 as network input, namely fusing the high-dimensional characteristics of the denoised images with the high-dimensional characteristics of the original images.
4. The SAR image oil spill detection method according to claim 3, wherein the specific flow of step 5 is as follows: 5.1, performing a convolution operation Ftr, giving an input , For the width and the height of the input matrix, C is the characteristic channel number of the input matrix, and simple convolution operation, namely a residual error network is performed; 5.2, performing Fsq extrusion operations, as shown in equation 2, the feature map is compressed in the direction along the spatial dimension, i.e., global average pooling is performed to obtain 1×1× Is a feature matrix of (1); (2) and 5.3, performing Fex excitation operation, and giving different weights W according to different feature importance, wherein the weights W are shown in a formula (3): (3) wherein: the character of the channel is described, Is used for reacting tensor The weight of the medium-feature map, The function is activated for the purpose of Swish, The function is activated for Sigmoid, , The weight for the full connection layer for dimension increasing and dimension decreasing; 5.4, final weight calibration operation, two-dimensional matrix Each value of (2) is multiplied by As shown in formula (4): (4) wherein: Is tensor The weight of the c-th feature map in (c), Index amount Feature mapping Is a product of the channels of (a).
5. The SAR image oil spill detection method according to claim 4, wherein the specific flow of step 6 is as follows: 6.1, performing a convolution operation Ftr, giving an input , C is the characteristic channel number of the input matrix for the width and the height of the input matrix, and simple convolution operation, namely residual network processing, is carried out; 6.2, performing Fsq extrusion operations, as shown in equations 5 and 6, the feature map is compressed in the horizontal and vertical directions, i.e., performing convolution operations with convolution kernels of1 XW and H1 to give H1X And 1 XWX Is a feature matrix of (1); (5) (6) And 6.3, performing Fex excitation operation, and giving different weights W according to different feature importance, wherein the weights W are shown in formulas (7) and (8): (7) (8) wherein: the character of the channel is described, And Is used for reacting tensor The weights of the middle-level and vertical direction feature maps, The function is activated for the purpose of Swish, The function is activated for Sigmoid, , The weight for the full connection layer for dimension increasing and dimension decreasing; and 6.4, final weight calibration operation.
6. The SAR image oil spill detection method according to claim 5, wherein the hole convolution in the structure of the golden sub-tower with hole convolution shown in step 7 is decomposed into two hole convolutions with different expansion ratios, the sensitivity to feature extraction is improved, and each layer is tightly connected to each other, so as to perform feature sharing.

Description

SAR image oil spill detection method Technical Field The invention relates to the field of synthetic aperture radar image oil spill segmentation, in particular to an SAR image oil spill detection method, which is based on coordination attention mechanism feature fusion U-Net. Background Pollution caused by petroleum leakage causes irreversible damage to the marine ecosystem. The oil spill region segmentation is a key step of SAR image oil spill detection. Synthetic Aperture Radar (SAR) has become an important technology for monitoring marine oil spill. Synthetic Aperture Radar (SAR) can provide electromagnetic information for detecting marine oil spills. SAR obtains electromagnetic information on the sea surface through a scattering mechanism. The scattering mechanism occurs on surfaces covered by oil slick and on clean sea surfaces, and the information obtained is different. For clean sea surfaces, strong bragg scattering occurs, which appears bright in SAR images. When oil spill occurs, it reduces bragg scattering, which appears dark in the SAR image. The traditional method has the unavoidable limitations that (1) the parameter selection of threshold segmentation has great influence on subjective factors or experiences, (2) single characteristic information cannot represent global characteristics and can influence the segmentation effect, and (3) the traditional method mainly adopts low-level information of images to complete segmentation tasks and is difficult to extract deep semantic information of the images. These limitations result in poor accuracy of the traditional method-based oil spill image segmentation. In recent years, some deep learning models are used for marine oil spill detection of SAR images, and deep learning is one branch of machine learning, aiming at solving the task of machine learning through neural network models. Unlike traditional machine learning algorithms, deep neural networks extract image features layer by layer through their deep network layers. Depth features are typically abstract and contain deep semantic information. In addition, the characteristic extraction process is automatic, manual participation is not needed, and the efficiency is greatly improved. Although the deep learning model obtains a better detection result in the oil spill detection task, some limitations still exist in further improving the detection accuracy. Due to the depth of the model, the feature extraction is insufficient, the receiving field is small, the loss of target information is caused, and redundant information can be generated by the model. Disclosure of Invention In order to solve the technical problems, the invention provides a SAR image oil spill detection method based on a feature fusion U-Net of a coordinated attention mechanism, which uses wavelet threshold transformation to remove image noise and performs feature fusion with an original image. By embedding the coordination attention module in the original network, high-level features in the image are extracted, redundant features are reduced, important features are emphasized, and the segmentation accuracy is improved. In order to achieve the above purpose, the present invention provides the following technical solutions: the SAR image oil spill detection method comprises the following steps: Step 1, 750 pictures are selected from European Space Agency (ESA) oil spill detection data sets to serve as training sets, 250 pictures are served as test sets, 90% of the training sets are randomly selected to serve as experimental training, and 10% of the training sets are used for experimental verification; Step 2, preprocessing the training set selected in the step 1, namely randomly cutting and splicing the training set, wherein the original SAR image format is 1250 multiplied by 650, and the size of the preprocessed data set picture is remolded into 256 multiplied by 256, and the number of characteristic channels is 3; step 3, denoising the training set obtained in the step 2 by using Wavelet Threshold Transformation (WTT); step 4, building a feature fusion U-Net model based on a coordinated attention mechanism, randomly selecting 50% training set pictures obtained in the step 2 and the step 3 as the input of a network model, and carrying out global feature extraction and fusion; step 5, using a Residual module (Residual Model), and adding an extrusion and excitation module (SE) after part of the Residual module to enable the Model to autonomously learn the weight coefficient of each channel; step 6, embedding a coordinated attention mechanism module (CA) at the jump joint to eliminate redundant information; Step 7, taking a space pyramid structure (ASPP) with cavity convolution as a bottom network, and extracting wider characteristic information while increasing a receptive field; and 8, inputting the selected test set into a trained feature fusion U-Net model based on a coordinated attention mechanism for testing, and obtaining a net