Search

CN-122023777-A - Ship small target detection method in SAR images with different qualities

CN122023777ACN 122023777 ACN122023777 ACN 122023777ACN-122023777-A

Abstract

The invention discloses a ship small target detection method in SAR images with different qualities, which is used for accurately detecting ship small targets in the SAR images with different qualities, and comprises the steps of synthesizing SAR images with actual SAR images and simulated speckle noise with various intensities under the guidance of SAR imaging mechanism, establishing synthetic SAR image data sets containing synthetic SAR images with different qualities, and carrying out ship small target detection on the synthetic SAR images, wherein the ship small target detection method comprises the steps of firstly realizing SAR speckle noise suppression based on a residual error network to obtain high-quality SAR images; the method comprises the steps of obtaining a ship small target edge feature, obtaining a SAR image with a prominent edge, obtaining a multi-level feature map based on a context enhancement strategy, obtaining a target detection result, obtaining a ship small target edge feature extraction based on a side window mean value filter, obtaining a SAR image with a prominent edge, obtaining a multi-level feature map fusion based on a feature purification treatment, and obtaining a target detection result. The method and the device ensure accurate detection of the small targets of the ship on the basis of adapting to SAR images with different qualities.

Inventors

  • ZHENG TONG
  • XU YEWANG

Assignees

  • 北京工商大学

Dates

Publication Date
20260512
Application Date
20260203

Claims (6)

  1. 1. The ship small target detection method in the SAR images with different qualities is characterized by comprising the following steps: constructing a synthetic SAR image dataset containing different qualities, wherein the different qualities are synthesized with a real SAR image by simulating SAR image speckle noise with different intensities; According to the constructed synthetic SAR image data set, establishing a residual network-based SAR image speckle noise suppression model to realize speckle noise suppression of synthetic SAR images with different qualities; according to the SAR image after the speckle noise suppression, establishing a ship small target edge feature extraction model based on side window mean filtering so as to enhance target edge feature expression; according to the SAR image with the protruding edge, a multi-level feature map extraction model based on a context enhancement strategy is established, multi-level expression of features is realized, and small target detection is assisted; and establishing a multi-level characteristic map fusion model based on characteristic purification treatment according to the extracted multi-level characteristic map, and fusing multi-level characteristic information to obtain a ship small target detection result.
  2. 2. The method for constructing a synthetic SAR image dataset of different quality according to claim 1, wherein the synthetic SAR image is simulated according to a SAR imaging mechanism, comprising: Deducing the statistical characteristics of speckle noise in the SAR image according to the SAR imaging mechanism, and generating simulated speckle noise with different intensities based on the statistical characteristics; deducing speckle noise in the SAR image to follow Gamma distribution according to the SAR imaging mechanism, combining the generated simulated speckle noise with different intensities with the real SAR image on the basis of the Gamma distribution to form synthetic SAR images with different qualities, and forming a data set.
  3. 3. The method for constructing the residual network-based SAR image speckle noise suppression model according to claim 1, wherein the design of the residual network structure and the loss function comprises: Based on convolution and activation processing, a single-layer noise suppression processing module is established, a plurality of modules are cascaded, a synthetic SAR image affected by speckle noise is input into the plurality of cascaded modules to obtain a speckle noise learning matrix, and the input synthetic SAR image and the speckle noise learning matrix are subjected to pixel-by-pixel division to obtain a SAR image after speckle noise suppression. When the speckle noise suppression model is trained, the synthetic SAR images with different qualities obtained in the claim 2 are used as the input of the model, the original real SAR image is used as a label, the loss calculation is carried out on the model output result and the label, and the parameter value in the model is adjusted through a counter-propagation mode, so that the model training is completed.
  4. 4. The method for constructing a multi-level feature map extraction model based on a context enhancement strategy as claimed in claim 1, wherein the design of the side window mean filter and the edge feature extraction comprises the following steps: According to the angular distribution of the ship target in the SAR image and the spindle-shaped shape characteristics, 12 groups of side window mean filters are designed to smooth speckle noise in the residual SAR image, and meanwhile target contour information is reserved. Comparing the result graph after 12 sets of side window mean value filtering with the SAR image after noise suppression, and selecting the result graph with 12 sets of results closest to the SAR image in each pixel as the value of the output result on the pixel. And respectively carrying out 4 groups of Sobel operator processing on the output results of the side window mean value filtering processing, extracting edge information in different directions, and then carrying out summation processing on the 4 groups of results to highlight the edge information of the SAR image in the 4 directions, namely an edge highlighting feature map. And (3) splicing the partial input SAR image with the edge salient feature image to obtain an edge salient SAR image.
  5. 5. The method for constructing the multi-level feature map extraction model based on the context enhancement strategy as claimed in claim 1, wherein the design of the context enhancement strategy and the multi-level feature map extraction comprises the following steps: The convolution, activation and pooling processing are carried out to form a single-layer context enhancement processing module, SAR images with prominent edges are processed by the processing module for 3 times, encoder feature images C3, C2 and C1 are sequentially obtained, the feature images are gradually reduced in size, and a multi-level feature image is obtained; And carrying out convolution processing with the same size as 3 convolution kernels on the encoder characteristic diagram C1 with the smallest size, and obtaining rich ship small target context semantic information through different convolution step sizes.
  6. 6. The method for constructing the multi-level feature map fusion model based on feature purification treatment according to claim 1, wherein the design of feature purification and multi-level feature map fusion comprises the following steps: Carrying out convolution processing with different convolution kernel sizes and the same convolution step length on 3 groups of feature images with different sizes obtained by the context enhancement processing to further obtain 3 groups of feature images with the same size, and then splicing the 3 groups of feature images to realize fusion of context semantic feature images; Carrying out convolution processing with the same convolution kernel size and the same step length on the encoder feature maps C3, C2 and C1 in claim 5, and carrying out point-by-point summation on the up-sampling processing results of the current layer feature map and the further layer feature map to obtain current layer feature maps in a decoding stage, wherein the current layer feature maps are F3, F2 and F1 respectively; And performing feature purification treatment on the output feature images F3, F2 and F1 in the decoding stage, performing up-sampling treatment and down-sampling treatment on the output feature images F3 and F1 respectively to ensure that the sizes of the output feature images are consistent with those of the output feature images F2, splicing the output feature images and the output feature images to obtain a feature image matrix, performing two paths of parallel treatment on the feature image matrix, wherein one path is activation treatment based on a Softmax function, the other path is respectively subjected to maximum pooling treatment and mean pooling treatment, and summing the feature image matrix and the feature image matrix after convolution, performing point-by-point multiplication on the summation result and the activation treatment result to obtain a feature image after fusion, filtering redundant and contradictory information of feature images with different sizes, and obtaining SAR image ship target detection results through a classical detection head.

Description

Ship small target detection method in SAR images with different qualities Technical Field The invention relates to the field of remote sensing imaging, in particular to SAR image target detection in remote sensing imaging. Background SAR has the advantages of all-weather, cloud penetration, fog penetration, long acting distance and the like, and is mainly used in civil fields such as military and topography mapping, disaster assessment, environment monitoring and the like, such as strategic tactical target detection, accurate guidance, accurate target striking and the like. SAR target detection is an important branch of SAR image interpretation. SAR image target detection has special difficulties. In the general field of target detection, when the number of the pixel points of the target is less than 32×32 or the product of the length and width of the target marking frame is divided by the product of the length and width of the whole image, and the result of opening the root number is less than 3%, the task belongs to a small target detection task. SAR image ship target detection accords with the description, and is generally expressed as a small target detection condition under a large scene. The target size is small, and the difficulty of target detection is increased. Secondly, inherent speckle noise in the SAR image can cause image quality reduction, and SAR image processing difficulty is increased. Speckle noise is caused during SAR image generation. Which takes the form of granular noise, severely affecting SAR image quality. In addition, the intensity of speckle noise is directly related to parameters of SAR imaging systems, so that it is difficult to ensure that SAR images with similar quality are obtained by different acquisition systems. This also puts higher demands on the adaptability of SAR image target detection. Furthermore, with the continued development of artificial intelligence technology, SAR image target detection based on artificial intelligence technology has taken a key position, and especially research based on convolutional neural networks (Convolutional Neural Network, CNN) has become a hotspot in the art. The invention provides a ship small target accurate detection method which can adapt to SAR images with different qualities by taking SAR imaging mechanism as theoretical guidance and artificial intelligence as technical support. Disclosure of Invention The embodiment of the invention provides a ship small target detection method in SAR images with different qualities, which is used for realizing accurate detection of the ship small target aiming at the SAR images with different qualities, and comprises the following steps: 1) Under the guidance of SAR imaging mechanism, synthesizing SAR images by using actual SAR images and various simulated speckle noise with different intensities, and taking the SAR images as training and testing data sets of a target detection model; 2) Establishing a speckle noise suppression model, taking a synthetic SAR image as input, learning speckle noise in the synthetic SAR image based on a residual network, and obtaining a SAR image after speckle noise suppression according to the multiplicative noise characteristic of the speckle noise; 3) Establishing a ship small target edge feature extraction model based on side window mean filtering, setting a plurality of side window mean filters, smoothing speckle noise remained in SAR images, retaining target contours, extracting edge features through a plurality of Sobel operators, and obtaining SAR images with prominent edge features after combining with input SAR images; 4) Establishing a multi-level feature map extraction model based on a context enhancement strategy, obtaining a deep feature map through multi-level convolution and activation processing, and extracting context semantic information in the deep feature map by using different convolution step sizes; 5) Establishing a multi-level feature map fusion model based on feature purification treatment, carrying out convolution treatment on semantic information of deep feature maps with different convolution kernel sizes, realizing fusion of context semantic information in a splicing mode after ensuring that the sizes are consistent, filtering redundant and contradictory information of feature maps with different sizes through feature purification treatment, and finally obtaining a final detection result through a detection head. Drawings FIG. 1 is a schematic diagram of a ship small target detection method in SAR images with different qualities in an embodiment of the invention; FIG. 2 is a schematic diagram of a method for creating a data set including synthetic SAR images of different levels according to an embodiment of the present invention; FIG. 3 is a schematic diagram of a method for establishing a speckle noise suppression model in an embodiment of the invention; FIG. 4 is a schematic diagram of a method for establishing a ship small target