Land SAR target detection method based on improved YOLOV5 network

Abstract

The invention discloses a land SAR target detection method based on an improved YOLOV5 network, and the method comprises the steps: 1, obtaining a remote sensing SAR image, and preprocessing the remote sensing SAR image into training data needed by network training; 2, improving a YOLOV5 network model: firstly, replacing a Focus module in a backbone network with a 7 * 7 convolution module to increase a model receptive field, and introducing a modified CBS module; secondly, the SPPF module is used to improve the detection speed under the condition that the effect of the SPPF module is the same as that of an original module; and finally, introducing an SIOU loss function, and redefining the loss function by using a vector angle between boundary regression. 3, inputting the training data into the improved YOLOV5 model for training; 4, performing target detection and identification on the remote sensing image by using the trained model; the method can effectively solve the problem that small target features are lost in the sampling process under land target features, and consequently small target missing detection is caused, and the reasoning accuracy can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a land SAR target detection method based on an improved YOLOV5 network. Background Art

[0002] Land remote sensing SAR target detection has always been a hot research direction in the field of remote sensing images. With the development of modern satellite observation technology, the field of land remote sensing SAR detection has gained great significance, and remote sensing technology is also booming, playing an important role in many civil fields, such as aviation supervision, port and shipping transportation, urban planning and layout, disaster monitoring and early warning, land and natural resources monitoring, and ocean observation and research.

[0003] At present, with the improvement of computer computing power, the land SAR target detection method based on deep learning has achieved great success in remote sensing image detection. Target detection is at the core of computer vision. With the development of convolution, target detection algorithms have made great progress. Representative two-stage algorithms include Fast-RCNN and Faster-RCNN. One-stage series algorithms have also made rapid progress in recent years, and representative algorithms include Yolo series.

[0004] In the complex background of land remote sensing SAR images, the characteristics of the images themselves include large background interference, large changes in the aspect ratio of the target size range, and a small number of pixels for small targets. This makes it difficult to extract features and distinguish between targets and backgrounds. In addition, in the commonly used deep learning detection network, the input image is downsampled by integer multiples to obtain a larger receptive field, which will cause the target features to be lost. The unique characteristics of land remote sensing SAR images also bring great difficulties and challenges to target detection. The characteristics of remote sensing SAR images and the challenges of remote sensing image target detection are summarized as follows:

[0005] ⑴ The land background is complex and the interference is large: With the improvement of resolution and the large imaging area, there are a lot of natural and artificial backgrounds in land remote sensing images, which seriously interfere with target detection.

[0006] (2) The target size range and aspect ratio vary greatly: The scales of land remote sensing images vary greatly in different scenes, and targets of the same category also have different scales.

[0007] ⑶ There are many small targets: Land aircraft and vehicle targets are small in size. In the existing detection methods, the features of small targets are easily lost due to feature downsampling, resulting in missed detection of small targets, which increases the difficulty of detection.

[0008] In summary, land target detection has problems such as large background interference, large changes in aspect ratio, and difficulty in detecting small targets. At the same time, the existing deep learning network technology has limited perception capabilities for small targets, resulting in insufficient detection precision and low accuracy for small targets such as vehicles. Summary of the invention

[0009] In view of this, the present invention provides a land SAR target detection method based on an improved YOLOV5 network, which can detect small land SAR targets and effectively improve the detection accuracy.

[0010] The technical solution for implementing the present invention is as follows:

[0011] A land SAR target detection method based on an improved YOLOV5 network comprises the following steps:

[0012] Step 1: Obtain remote sensing SAR images and preprocess them into training data required for network training;

[0013] Step 2: Improve the YOLOV5 network model: First, replace the Focus module in the backbone network with a 7*7 convolution module to increase the model receptive field, and introduce the modified CBS module; secondly, use the SPPF module to improve the detection speed while ensuring the same effect as the original module; finally, introduce the SIOU loss function and redefine the loss function using the vector angle between boundary regressions;

[0014] Step 3: Input the training data into the improved YOLOV5 model for training;

[0015] Step 4: Use the trained model to detect and identify targets in remote sensing images.

[0016] Furthermore, step one specifically includes:

[0017] S11: Download the corresponding land remote sensing SAR dataset from the database for preprocessing, select images containing target information for annotation, convert them into XML file format after annotation, and describe the target frame with 8 parameters (x1, y1, x2, y2, x3, y3, x4, y4); where x1 is the upper left corner horizontal coordinate, y1 is the upper left corner vertical coordinate, x2 is the lower left corner horizontal coordinate, y2 is the lower left corner vertical coordinate, x3 is the upper right corner horizontal coordinate, y3 is the upper right corner vertical coordinate, x4 is the lower right corner horizontal coordinate, and y4 is the lower right corner vertical coordinate;

[0018] S12: Convert the 8-parameter target box into a 5-parameter target box (tx, ty, tw, th, tn), where tx, ty are the horizontal and vertical coordinates of the center of the target box, tw, th are the width and height of the target box, and tn represents the category;

[0019] S13: Divide each type of sample into a training sample set and a validation sample set in a ratio of 7:3, and shuffle and package the two into the data format required by the network.

[0020] Furthermore, in step 2, the Focus structure in the Backbone network is replaced with a convolution model CBS with a convolution kernel size of 7, and the convolution in CBS is replaced with a deformable convolution Dconv.

[0021] Furthermore, in step 2, an SPPF module is used in the Backbone network to replace the SPP module. The SPPF module replaces the three parallel maximum pooling layers with convolution kernels of 5, 9, and 13 in the SPP module with three serial maximum pooling layers with a convolution kernel of 5.

[0022] Furthermore, in step 2, SIOU Loss is used as the positioning loss function. The SIOU loss function consists of Anglecost, Distance cost, Shape cost, and IOU cos. SIOU takes into account the Angle cost between regressions and redefines the loss function.

[0023] Beneficial effects:

[0024] 1. The present invention proposes an improvement on the YOLOV5 Backbone network and loss function, which can effectively address the problem of missing small target features during the downsampling of land target features, resulting in missed detection of small targets, and can effectively improve the accuracy of reasoning.

[0025] 2. The present invention replaces the Focus structure in the Backbone network with a convolution model CBS with a convolution kernel size of 7, and replaces the C in CBS with a deformable convolution Dconv, so that the sampling position is more consistent with the shape and size of the object itself. Compared with the original convolution, the deformable convolution can better eliminate the interference of background noise, better consider the changes in the object shape, adaptively adjust, and finally perform downsampling.

[0026] 3. The present invention adopts the SPPF module to replace the SPP module. The improved module improves the model reasoning speed while ensuring the same effect as the original structure.

[0027] 4. The present invention replaces CIOU Loss with SIOU Loss as the positioning loss function. SIOU takes into account the Angle cost between regressions, redefines the loss function, and effectively improves the accuracy of reasoning. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a flow chart of the method of the present invention.

[0029] Figure 2 This is the Yolov5 structure diagram of the present invention.

[0030] Figure 3 It is the SPPF structure diagram of the present invention.

[0031] Figure 4 These are the detection effect diagrams of the model of the present invention; (a) is the vehicle detection effect diagram, and (b) is the aircraft detection effect diagram. DETAILED DESCRIPTION

[0032] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0033] The present invention provides a terrestrial SAR target detection method based on an improved YOLOV5 network, such as Figure 1 As shown, the specific steps include:

[0034] S1: Acquire remote sensing SAR images;

[0035] S11: Download the corresponding land remote sensing SAR data set from the database for preprocessing, select images containing target information for annotation, convert them into XML file format after annotation, and describe the target frame with 8 parameters (x1, y1, x2, y2, x3, y3, x4, y4). Among them, x1 is the upper left corner horizontal coordinate, y1 is the upper left corner vertical coordinate, x2 is the lower left corner horizontal coordinate, y2 is the lower left corner vertical coordinate, x3 is the upper right corner horizontal coordinate, y3 is the upper right corner vertical coordinate, x4 is the lower right corner horizontal coordinate, and y4 is the lower right corner vertical coordinate;

[0036] S12: Convert the 8-parameter target box into a 5-parameter target box (tx, ty, tw, th, tn), where tx, ty are the horizontal and vertical coordinates of the center of the target box, tw, th are the width and height of the target box, and tn represents the category;

[0037] S13: Divide each type of sample into a training sample set and a validation sample set in a ratio of 7:3, and shuffle and package the two into the data format required by the network.

[0038] S2: Improve the YOLOV5 model to obtain an improved YOLOV5 model;

[0039] S21: Since land images have large sizes and different shapes, the original YOLOV5 backbone network uses the Focus structure for downsampling. The wide and high dimensional information of the land is converted to the channel dimension through slicing, and then the target features are extracted through a 3*3 convolution kernel. With this structure, when the input image size is large, the network initially uses a smaller convolution kernel, which will lose the overall features of the large target. Therefore, the improved method is to replace the Focus structure in the Backbone network with a modified convolution model CBS with a convolution kernel size of 7, where C in CBS is conv convolution, B is BN normalization, and S is sliu activation function. Here, C in CBS is replaced by deformable convolution Dconv (Deformable Conv). Because the remote sensing image data of the present invention is relatively complex, the convolution kernel of conventional convolution is of fixed size and shape, which may have better effects on objects with regular shapes. If irregularly shaped objects or relatively complex objects are encountered, the detection effect will be affected. After the convolution is changed to Dconv, its sampling position is more in line with the shape and size of the object itself, while the standard convolution form cannot do this. Compared with the original convolution, the deformable convolution can better eliminate the interference of background noise, can better consider the changes in the object's body shape, adaptively adjust, and finally downsample.

[0040] S22: In the Backbone network, the SPPF structure is used instead of the SPP structure. The SPPF structure is as follows: Figure 3 As shown in the figure, the SPPF module replaces the three parallel maximum pooling layers with convolution kernels of 5, 9, and 13 in the SPP module with three serial maximum pooling layers with a convolution kernel of 5. The improved module improves the model inference speed while ensuring the same effect as the original structure.

[0041] S23: YOLOV5 has three loss functions, namely classification loss, confidence loss and localization loss. The localization loss function uses CIOU Loss as the loss function. The CIOU Loss formula is as follows.

[0042]

[0043] Among them, b and b gt Represent the center points of the prediction box and the GT box respectively, p(b,b gt ) represents the Euclidean distance between the center points of the prediction box and the GT box, c is the diagonal length of the minimum circumscribed rectangle of the prediction box and the GT box, a is the weight function, v is a measure of the consistency of the aspect ratio between the two boxes, and w gt / h gtThey represent the aspect ratios of the GT box and the predicted box respectively.

[0044] CIOU Loss takes into account the overlapping area between the prediction box and the target box, and the aspect ratio of the center point distance, but does not take into account the mismatch angle problem between the real box and the prediction box. Therefore, the present invention introduces SIOU Loss as the positioning loss function of Bounding box regression. The SIOU loss function consists of Angle cost, Distance cost, Shape cost, and IOU cos. The SIOU Loss regression loss function formula is shown below.

[0045]

[0046]

[0047] In the formula, Λ is the Angle cost, △ is the Distance cost under the condition of Angle cost, and Ω is the Shape cost. SIOU takes into account the Angle cost between regressions, redefines the loss function, and effectively improves the accuracy of reasoning. Angle cost describes the minimum angle between the center point connection and the xy axis. When the center point is aligned on the x-axis or y-axis, Λ=0. When the center point is connected to the x-axis at 45°, Λ=1. This penalty can guide the anchor box to move to the nearest axis of the target box, reducing the total number of degrees of freedom of BBR. Distance cost describes the distance between the center points. The aspect ratio between the two boxes considered by Shape cost is defined by calculating the difference in width between the two boxes and the ratio of the maximum width between the two boxes (the same applies to length). The general idea is similar to CIOU, except that CIOU can consider the convergence of the overall shape of the two boxes, while SIoU achieves the effect of overall shape convergence by converging the length and width sides.

[0048] YOLOV5 is a classic single-stage target detection network, which mainly consists of three parts: input end, Backbone (backbone network), Neck (neck network), and Head detection head. The network structure is as follows: Figure 2 As shown in the figure. The input end is the image preprocessing part, which includes Mosaic data enhancement, adaptive anchor box calculation, and adaptive image scaling; Backbone is the feature extraction part, which includes Fcous structure, CSP_1X structure, and SPP structure; Neck is the feature fusion part, which includes CSP2_X structure and FPn+PAN structure. The Head detection head is the target result output, which includes CIOU_Loss and NMS non-maximum suppression.

[0049] S3: Input the processed land remote sensing data into the modified YOLOV5 network model for training;

[0050] S4: After the model training is completed, target detection and recognition are performed on the remote sensing image according to the trained model, the detection result is matched to the original image and the location information of the target is framed as the final detection result, and the detection result is saved or output. Figure 4 The following are model detection renderings, where (a) is a vehicle detection rendering and (b) is an aircraft detection rendering.

[0051] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A land SAR target detection method based on an improved YOLOV5 network, characterized in that: The following steps are involved: Step 1: Obtain remote sensing SAR images and preprocess them into training data required for network training; Step 2: Improve the YOLOV5 network model: First, replace the Focus module in the backbone network with a 7*7 convolution module to increase the model receptive field, and introduce the modified CBS module; secondly, use the SPPF module to improve the detection speed while ensuring the same effect as the original module; finally, introduce the SIOU loss function and redefine the loss function using the vector angle between boundary regressions; Step 3: Input the training data into the improved YOLOV5 model for training; Step 4: Use the trained model to detect and identify targets in remote sensing images.

2. The terrestrial SAR target detection method according to claim 1, characterized in that: Step 1 specifically includes: S11: Download the corresponding land remote sensing SAR dataset from the database for preprocessing, select images containing target information for annotation, convert them into XML file format and save them in (x1, y1, x2, y2, x3, y3, x4, y4) 8 parameters describe the target frame; x1 is the upper left corner horizontal coordinate, y1 is the upper left corner vertical coordinate, x2 is the lower left corner horizontal coordinate, y2 is the lower left corner vertical coordinate, x3 is the upper right corner horizontal coordinate, y3 is the upper right corner vertical coordinate, x4 is the lower right corner horizontal coordinate, y4 is the lower right corner vertical coordinate; S12: Convert the 8-parameter target box into a 5-parameter target box (tx, ty, tw, th, tn), where tx, ty are the horizontal and vertical coordinates of the center of the target box, tw, th are the width and height of the target box, and tn represents the category; S13: Divide each type of sample into a training sample set and a validation sample set in a ratio of 7:3, and shuffle and package the two into the data format required by the network.

3. The terrestrial SAR target detection method according to claim 1, characterized in that: In step 2, the Focus structure in the Backbone network is replaced with a convolution model CBS with a convolution kernel size of 7, and the convolution in CBS is replaced with a deformable convolution Dconv.

4. The terrestrial SAR target detection method according to claim 3, characterized in that: In step 2, the SPPF module is used in the Backbone network to replace the SPP module. The SPPF module replaces the three parallel maximum pooling layers with convolution kernels of 5, 9, and 13 in the SPP module with three serial maximum pooling layers with a convolution kernel of 5.

5. The terrestrial SAR target detection method according to claim 3 or 4, characterized in that: In step 2, SIOULoss is used as the positioning loss function. The SIOU loss function consists of Angle cost, Distance cost, Shape cost, and IOUcos. SIOU takes into account the Angle cost between regressions and redefines the loss function.

Images