A SAR image aircraft detection method based on attention enhanced network

Abstract

The application provides a SAR image airplane detection method based on an attention enhancement network, small-scene SAR images are obtained by cutting a large-scene SAR image and are subjected to data enhancement preprocessing; the small-scene SAR images are input into an improved target detection network to detect airplane targets; false alarm targets are filtered out by inputting the detected airplane targets into an image post-processing network, and finally, airplane target detection results are obtained. Compared with a traditional SAR image airplane detection algorithm, the scheme reduces the calculation complexity, reduces the calculation memory requirement and increases the sample diversity through the cutting and data enhancement operations in the preprocessing, and improves the airplane target detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of synthetic aperture radar target detection, and particularly relates to a SAR image airplane detection method based on an attention enhanced network. BACKGROUND

[0002] Synthetic aperture radar (SAR) is widely used in military and civilian fields due to its all-weather and all-day observation capability, and the corresponding interpretation technology has also developed rapidly. The airplane target has the characteristics of high value and time sensitivity, and accurate and rapid detection of the airplane target can provide strong information support for military decision-making.

[0003] Although SAR images are widely used for detecting airplane targets, there are still some problems to be solved. First, the detection efficiency of the algorithm for airplane targets is low. The airplane target occupies a small proportion in a large scene SAR image, and if the whole image is traversed to search for the airplane target, the running efficiency of the algorithm will be inevitably reduced. Second, the feature extraction capability of the algorithm for airplane targets is weak. The airplane target often appears as a discrete scattering point distribution in the SAR image, is sensitive to attitude, and has coherent speckle noise. How to effectively represent the airplane target in a complex scene is a difficulty in SAR image airplane target detection. Finally, the detection accuracy of the algorithm for airplane targets is low. The airplane target in the SAR image is often affected by other surrounding ground objects, resulting in a high false alarm rate. The high false alarm rate will reduce the credibility of the algorithm detection result and has no practical application value. In addition, missing or misidentifying the target will lead to the loss of key targets.

[0004] In summary, the SAR image airplane target detection has the problems of low efficiency, poor real-time performance, low detection accuracy, high false alarm rate and missing detection. With the increasing amount of SAR data obtained by a large number of orbiting radar satellites every day, the demand for high-efficiency and high-precision airplane target detection algorithms has become increasingly urgent. SUMMARY

[0005] The SAR image airplane detection method based on the attention enhanced network provided by the present application mainly solves the technical problem of how to improve the detection accuracy of the SAR image airplane target.

[0006] To solve the above technical problems, the present application provides a SAR image airplane detection method based on an attention enhanced network, which comprises the following steps:

[0007] The original large scene SAR image is cut to obtain a small scene SAR image, and the small scene SAR image is preprocessed by data enhancement;

[0008] The preprocessed small scene SAR image is input into an improved target detection network to detect the airplane target.

[0009] According to the detected aircraft target, the image post-processing network is sent to filter false alarm targets, and final aircraft target detection results are obtained.

[0010] Optionally, the SAR image cutting operation refers to changing a large scene SAR image into a uniform small scene SAR image.

[0011] Optionally, the data enhancement preprocessing includes at least one of the following operations: random rotation, mosaic enhancement, and test data enhancement (TTA).

[0012] Optionally, the improved target detection network is an improved Yolov5 single-stage detection network, the original backbone convolution layer in Yolov5 is replaced by a Swin-Transformer attention network for extracting effective features of an aircraft target, and the original loss function part C-IoU Loss in Yolov5 is replaced by an alpha-IoU Loss for aircraft target bounding box regression and target detection.

[0013] Optionally, the image post-processing network is a weighted bounding box fusion (WBF) network for filtering redundant false alarm targets detected for the same aircraft target, and finally splicing the detected aircraft targets in all small scene SAR images into a large scene SAR image.

[0014] The present application has the following advantages:

[0015] According to the SAR image aircraft detection method based on the attention enhanced network, the small scene SAR image is obtained by cutting the large scene SAR image and is subjected to data enhancement preprocessing, the small scene SAR image is input into the improved target detection network to detect the aircraft target, and according to the detected aircraft target, the image post-processing network is sent to filter false alarm targets, and final aircraft target detection results are obtained. Compared with the traditional SAR image aircraft detection algorithm, the present application reduces the calculation complexity, reduces the calculation memory requirement and increases the sample diversity through the cutting and data enhancement operation in the preprocessing, and improves the aircraft target detection precision.

[0016] Further, the target detection network of the present application adopts the single-stage target detection Yolov5 network improved on the backbone convolution layer and the loss function, which can maintain high detection efficiency and high detection precision, the introduction of the Swin-Transformer attention network can allocate resources according to the weight of each part of the SAR image, and better extract the effective features of the aircraft in the SAR image, and the introduction of the loss function alpha-IoU Loss can obtain more accurate aircraft target bounding box regression and target detection.

[0017] Further, the scheme can filter out the common redundant false alarm targets in the SAR image by introducing a weighted bounding box fusion (WBF) network, and improve the final detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of the SAR image aircraft target detection method of the present application is shown in the figure.

[0019] Figure 2 The specific structure diagram of the attention network Swin-Transformer of the present application is shown in the figure. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the specific embodiments and the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0021] Embodiment one:

[0022] The present embodiment provides a SAR image aircraft detection method based on attention enhanced network, and the flowchart is shown in the figure. Figure 1

[0023] First, the original large scene SAR image is cut into uniform small scene SAR images, and specifically, the cutting can be performed according to the set small scene size. The small scene SAR images obtained are preprocessed by data enhancement, including but not limited to random rotation, mosaic enhancement (mosaic), test data enhancement (TTA) and the like, to increase sample diversity and improve aircraft target detection accuracy. Then, the preprocessed small scene SAR images are input into the improved target detection Yolov5 network to detect aircraft targets. Finally, the detected aircraft targets are sent to the image post-processing weighted bounding box fusion (WBF) network, and the redundant false alarm targets detected for the same aircraft target are filtered out. Finally, all the aircraft targets detected in the small scene SAR images are spliced back into the large scene SAR image, which is the final aircraft target detection result.

[0024] The specific implementation process of the present scheme is as follows:

[0025] Step one, cutting the original large scene SAR image and performing data enhancement preprocessing

[0026] ​The original large scene SAR image is cut into uniform small scene SAR images, and data enhancement operations such as random rotation, mosaic enhancement (mosaic), test data enhancement (TTA) and the like are performed on the small scene SAR images, so as to reduce the image processing calculation amount and increase the sample diversity. Among them, according to experience, when the image random rotation data enhancement operation is rotated by 5 degrees, the effective features of the aircraft target in the SAR image are basically unchanged, and the detection effect is the best. The present application mainly improves the target detection Yolov5 network and the weighted boundary box fusion (WBF) network in image post-processing, so as to improve the detection accuracy of the aircraft target in the SAR image.

[0027] Step two, improved target detection Yolov5 network

[0028] According to the improved target detection network, the aircraft target is detected by inputting the small scene SAR image after preprocessing. Since the target detection Yolov5 network except the backbone convolution layer and the loss function part is not the focus of the present application, it will not be repeated here. The focus of the present application is the introduction of attention network Swin-Transformer and loss function α-IoU Loss in Yolov5, including the following steps:

[0029] a1. Extracting effective features of aircraft by attention network Swin-Transformer

[0030] The attention network Swin-Transformer replaces the original Yolov5 backbone convolution layer (Backbone), which can better extract effective features of aircraft in SAR images. The attention network Swin-Transformer constructs hierarchical Transformer by introducing the hierarchical construction method commonly used in traditional convolutional neural network (CNN) and introduces the local (locality) idea to calculate self-attention in non-overlapping window area, solving the scale and resolution problem of attention network Transformer migrated to computer vision (CV).

[0031] Please refer to Figure 2 The specific structure of the attention network Swin-Transformer is as follows:

[0032] As can be seen from Figure (a), the basic flow of the attention network Swin-Transformer architecture is very similar to the convolutional neural network (CNN) architecture, and 4 stages are constructed in each stage, which are similar to the repeated units. First, the picture is input into the Patch Partition module for blocking, that is, each 4x4 adjacent pixels is a Patch, and then it is flattened in the channel direction.

[0033] Assuming that the input is an RGB three-channel picture, each patch has 4x4 = 16 pixels, and each pixel has R, G, B three values, so after flattening, there are 16x3 = 48 values, so after the Patch Partition, the image shape changes from [H, W, 3] to [H / 4, W / 4, 48].

[0034] Then, the channel data of each pixel is linearly transformed by the Linear Embeding layer, from 48 to C, that is, the image shape changes from [H / 4, W / 4, 48] to [H / 4, W / 4, C].

[0035] Then, four stages are constructed to construct feature maps of different sizes. Except for the first stage1, which first passes through a Linear Embeding layer, the remaining three stages are first passed through a Patch Merging layer for down-sampling. Then, the Swin Transformer Block is repeatedly stacked, and the number of stacked Swin Transformer Block is even.

[0036] It should be noted that the Block includes two structures, as shown in Figure 2 (b), the difference between the two structures is only that one uses a W-MSA structure and one uses a SW-MSA structure. Moreover, the two structures are used in pairs, first using a W-MSA structure, and then using a SW-MSA structure. Finally, for the classification network, a LayerNorm layer, a global pooling layer and a fully connected layer are connected to obtain the final output.

[0037] a2. Introduction of loss function α-IoU Loss

[0038] Bounding box regression locates the target in the image by predicting the bbox of the target, which is the basis of target detection. The original loss function part C-IoU Loss in Yolov5 is replaced by α-IoU Loss to obtain more accurate aircraft target bounding box regression and target detection. The introduction of the new loss function α-IoU can significantly exceed the existing IoU-based loss; by adjusting the parameter α, the detector has greater flexibility in achieving different levels of bbox regression accuracy; it is more robust to small data sets and noise.

[0039] The specific implementation of the loss function α-IoU Loss is as follows:

[0040] First, apply the Box-Cox transformation to the IoU loss L IoU = 1-IoU, and generalize it to power IoU loss, denoted as α-IoU, where α>0, the formula is as follows:

[0041] L α-IoU = (1-IoU α ) = α

[0042] Here α-IoU is further simplified to

[0043] L α-IoU = 1-IoU α

[0044] and generalize it to a more general form by adding an additional power regularization term. This enables the proposed loss function α-IoU to generalize the existing IoU-based losses, including GIoU, DIoU and CIoU, into a new power IoU loss function, to obtain more accurate bounding box regression and target detection. Relative to L IoU , L α-IoU (α>1) increases the loss and gradient of high IoU targets, thereby improving the bbox regression accuracy. Empirically, α is not overly sensitive to different models or data sets, and in most cases, α=3 performs consistently well. The α-IoU loss can be easily used to improve the effectiveness of the detector, without introducing additional parameters or increasing training / inference time in clean or noisy environments.

[0045] Step three, weighted bounding box fusion (WBF) network filters out redundant false alarm targets

[0046] The final detected SAR image aircraft target may be selected by multiple target boxes, so the weighted bounding box fusion (WBF) network in the image post-processing network is used to filter the redundant false alarm targets in the output results of the improved target detection Yolov5 network, which can effectively improve the final detection accuracy. Finally, all the aircraft targets detected in the small scene SAR image are spliced back into the large scene SAR image.

[0047] The weighted bounding box fusion (WBF) network is implemented as follows:

[0048] Each predicted box of the model is added to a single list B. The list is sorted in descending order of confidence C; declare empty lists L and F for box clusters and fusion boxes, respectively. Each position in list L can contain a set of boxes (or a single box), forming a cluster; each position in F contains only one box, which is the fusion box from the corresponding cluster in L. The equation for generating the fusion box will be discussed later; loop through the predicted boxes in B, try to find a matching box in list F. Matching is defined as a box that overlaps the question box (IoU>THR) to a large extent; if no match is found, add the box in list B to the end of list L and F as a new entry; continue to the next box in list B; if a match is found, add this box to list L at the position pos corresponding to the matching box in list F; recalculate the box coordinates and confidence score in F[pos] using all T boxes accumulated in cluster L[pos], and the fusion formula is as follows:

[0049]

[0050]

[0051]

[0052] Obviously, those skilled in the art should understand that the steps of the present application described above can be realized by general computing devices, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by computing devices, so that they can be stored in computer storage media (ROM / RAM, magnetic disc, optical disc) and executed by computing devices, and in some cases, the steps shown or described can be executed in different order, or they can be made into individual integrated circuit modules, or multiple modules or steps among them can be made into a single integrated circuit module. Therefore, the present application is not limited to any specific combination of hardware and software.

[0053] The above is further detailed description of the present application in combination with specific embodiments, and cannot be deemed as limitation of the specific embodiments of the present application. For those skilled in the art of the present application, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which shall be deemed as falling within the protection scope of the present application.

Claims

1. A method for aircraft detection in SAR images based on attention-enhanced networks, characterized in that: include: The original large-scene SAR image is cut to obtain a small-scene SAR image, and the small-scene SAR image is preprocessed with data enhancement; The pre-processed small scene SAR image is input into the improved target detection network to detect the aircraft target; The detected aircraft targets are sent to the image post-processing network to filter out false alarm targets and obtain the final aircraft target detection results; The image post-processing network is a weighted bounding box fusion (WBF) network, which is used to filter out redundant false alarm targets detected for the same aircraft target, and finally stitch all aircraft targets detected in the small scene SAR image into the large scene SAR image; The SAR image cutting operation refers to converting a large scene SAR image into a uniform small scene SAR image; The data enhancement preprocessing includes at least one of the following operations: random rotation, mosaic enhancement, and test data augmentation (TTA); The improved target detection network is an improved Yolov5 single-stage detection network, in which the main convolutional layer in the original Yolov5 is replaced by a Swin-Transformer attention network to extract effective features of aircraft targets; The C-IoU Loss loss function in the original Yolov5 is replaced by α-IoU Loss for aircraft target bounding box regression and target detection.

Images