An unmanned aerial vehicle small target recognition method based on wavelet decomposition and motion vector

Abstract

The application relates to the technical field of target recognition, and specifically provides a small target recognition method for an unmanned aerial vehicle based on wavelet decomposition and a motion vector, which comprises the following steps: performing three-level wavelet decomposition on a small target image to obtain a third-layer low-frequency information region and a plurality of high-frequency detail information regions; performing image enhancement processing on the plurality of high-frequency detail information regions; performing image reconstruction on the plurality of high-frequency detail information regions after image enhancement and the third-layer low-frequency information region to obtain a reconstructed image; inputting the reconstructed image into a YOLOv5 improved model to recognize small targets in the reconstructed image; and performing false detection judgment on the recognized small targets. The application can effectively improve the recognition precision of small targets in a complex background by using wavelet transformation to enhance image detail information, and can further improve the precision of a target recognition algorithm by removing false detection caused by part of environmental noise through false detection judgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target recognition, and specifically provides a small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vectors. BACKGROUND

[0002] In recent years, unmanned aerial vehicles have been widely used in remote sensing surveying and mapping, resource exploration, post-disaster search and rescue, border patrol, material transportation, information acquisition and other fields, and a large number of “black flight” and “random flight” situations have occurred, posing a great threat to social safety. Research on small target detection and recognition of unmanned aerial vehicles has become an urgent practical demand.

[0003] In the field of small target recognition, the early research focus was mainly on the direction based on traditional image processing and pattern recognition technology. For example, methods such as template matching, edge detection, and feature extraction (such as Haar features, HOG features, etc.). At present, the above-mentioned small target recognition has achieved certain results in some simple scenarios, but in complex environmental conditions, the above-mentioned methods have many limitations.

[0004] With the development of machine learning technology, researchers have begun to try to combine traditional image processing methods with machine learning algorithms, such as support vector machines (SVM) and random forests (Random Forests) methods, which are widely used in small target recognition. By training a large number of positive and negative samples, the accuracy of machine learning algorithms in identifying small targets can be effectively improved, and they can better adapt to background changes. In recent years, the rise of deep learning, especially convolutional neural networks (CNN), has greatly promoted the development of small target recognition technology. Based on CNN, small target recognition methods can automatically extract more discriminative features from images using the powerful feature learning ability of CNN, thereby improving the accuracy of small target recognition in complex backgrounds. Target detection algorithms such as YOLO (You Only Look Once) and Faster R-CNN have high real-time performance and accuracy, and have been widely used in small target recognition of unmanned aerial vehicles.

[0005] The YOLO series model is a relatively excellent model in target detection. Although the YOLO series model, like most other target detection models, originated from AlexNet, in terms of accuracy, the YOLO series model is not inferior to the accuracy of the two-stage detector R-CNN, Fast R-CNN and Faster R-CNN. At the same time, the YOLO series model can maintain a high frame rate in real-time monitoring, so the YOLO series model is one of the most widely used models at present.

[0006] Although in recent years, deep learning has developed rapidly, the target detection field has made significant progress, especially the application of methods based on convolutional neural network (CNN) and region convolutional neural network (R-CNN), many methods perform well in large-scale target detection, but there are still significant deficiencies in small target detection. In the field of small target detection, deep learning still faces many limitations that need to be broken through. Traditional target detection algorithms have inherent shortcomings in feature extraction mechanisms, and their design logic is mostly based on large receptive fields to capture global features. This pattern can effectively handle regular size targets, but it inevitably weakens the ability to capture detailed information, resulting in a serious lack of feature representation for small-scale targets. At the same time, existing small target datasets for unmanned aerial vehicles are usually scarce, which leads to insufficient recognition ability of the trained model for small targets and easy deviation.

[0007] How to improve the accuracy and stability of small target recognition in complex background is still a difficult problem to be solved. SUMMARY

[0008] To solve the above problems, the present application provides a small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vector, which performs three-level wavelet decomposition on small target images, completes image enhancement and image reconstruction, uses a YOLOv5 improved model to recognize small targets in the reconstructed image and performs false detection judgment, greatly improving the small target recognition accuracy.

[0009] The small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vector provided by the present application comprises:

[0010] S1: performing three-level wavelet decomposition on small target images to obtain a third layer low-frequency information region and a plurality of high-frequency detail information regions;

[0011] S2: performing image enhancement processing on the plurality of high-frequency detail information regions, and performing image reconstruction on the plurality of high-frequency detail information regions after image enhancement and the third layer low-frequency information region to obtain a reconstructed image;

[0012] S3: improving the YOLOv5 model using a CBAM attention module and training the improved YOLOv5 model to obtain a YOLOv5 improved model;

[0013] S4: using the YOLOv5 improved model to recognize small targets in the reconstructed image;

[0014] S5: obtaining detection box position information corresponding to a plurality of consecutive reconstructed images, calculating the motion vector of the small target according to the detection box position information, calculating the acceleration module length according to the motion vector, and judging whether the recognized small target is a real target according to the acceleration module length.

[0015] Preferably, the three-stage wavelet decomposition process is: performing first-stage wavelet decomposition on the small target image to obtain a first layer low-frequency information region; performing second-stage wavelet decomposition on the first layer low-frequency information region to obtain a second layer low-frequency information region; and performing third-stage wavelet decomposition on the second layer low-frequency information region to obtain a third layer low-frequency information region.

[0016] Preferably, the image enhancement is realized through gamma transformation.

[0017] Preferably, the method for improving the YOLOv5 model by using the CBAM attention module is inserting the CBAM attention module into the backbone of the YOLOv5 model, and the specific insertion position is the first C3 module after input.

[0018] Preferably, the training method of the improved YOLOv5 model is: obtaining an existing small target image, obtaining a reconstructed image corresponding to the existing small target image through the method of steps S1 and S2, and training the improved YOLOv5 model by using the reconstructed image corresponding to the existing small target image.

[0019] Preferably, the method for calculating the acceleration module length is: calculating a velocity vector according to adjacent detection frame position information, calculating an acceleration vector according to adjacent velocity vectors, and calculating the acceleration module length according to the acceleration vector.

[0020] Preferably, the method for judging the real target is: setting a stability threshold , if the sample variance of the acceleration module length is not greater than the stability threshold , the corresponding target is a real target.

[0021] Preferably, the sample variance of the acceleration module length is:

[0022] ;

[0023] Wherein, , and respectively represent the first acceleration module length, the second acceleration module length and the third acceleration module length, represents the average value of the acceleration module length.

[0024] Preferably, the number of continuous multiple frames of reconstructed images is five frames.

[0025] Compared with the prior art, the present application can achieve the following beneficial effects:

[0026] By wavelet decomposition and image reconstruction, the image enhancement processing is carried out on the high-frequency detail information area, the edge and texture features of small targets can be effectively highlighted, the contrast between small targets and background is improved, the background interference is reduced, and the small targets are more easily detected; the CBAM attention module is introduced into the YOLOv5 model, the advantages of the YOLOv5 model are retained, the importance of different positions and channels in the small target image can be adaptively learned, important features are assigned with greater weights, and unimportant features are suppressed, so that the model can pay more attention to the key features related to the small targets, which is helpful to improve the feature extraction capability of the YOLOv5 model for small targets and enhance the recognition precision and robustness of the YOLOv5 model for small targets; meanwhile, the misjudgment of the recognized small targets is judged, the position information continuity of the detection frame is combined, the static background interference is filtered, and the recognition stability in the dynamic scene is improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 It is a flow chart of a small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vectors according to an embodiment of the application;

[0028] Figure 2 It is a flow chart of wavelet decomposition and image reconstruction according to an embodiment of the application;

[0029] Figure 3 It is a structure diagram of a YOLOv5 model according to an embodiment of the application;

[0030] Figure 4 It is a structure diagram of a CBAM attention module according to an embodiment of the application;

[0031] Figure 5 It is a schematic diagram of small targets recognized by the small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vectors according to an embodiment of the application;

[0032] Figure 6 It is a simulation diagram of detection frame position information according to an embodiment of the application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not constitute a limitation on the present application. In different embodiments, similar elements are associated with similar element labels. In the following embodiments, many details are described in order to make the present application better understood. However, those skilled in the art can easily recognize that some features can be omitted in different cases, or can be replaced by other elements, materials or methods. In some cases, some operations related to the present application are not shown or described in the specification in order to avoid the core part of the present application being overwhelmed by too much description, and it is not necessary for those skilled in the art to describe these related operations in detail according to the description in the specification and general technical knowledge in the art.

[0034] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other to form various embodiments without conflict. Meanwhile, each step or action in the method description can be sequentially adjusted or adjusted in a manner that can be easily seen by those skilled in the art. Therefore, the various sequences in the specification and drawings are only for the purpose of clearly describing a certain embodiment, and do not mean a necessary sequence, unless otherwise stated that a certain sequence must be followed.

[0035] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element indicated must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" and the like are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features limited by "first", "second" and the like can explicitly or implicitly include one or more features. In the description of the present application, unless otherwise stated, the meaning of "a plurality of" is two or more.

[0036] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0037] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] This invention provides a method for identifying small targets in unmanned aerial vehicles (UAVs) based on wavelet decomposition and motion vectors, used to identify UAVs in acquired small target images, such as... Figure 1 As shown, the specific details of the identification method are as follows:

[0039] S1: Perform three-level wavelet decomposition on the small target image to obtain the third-level low-frequency information region and multiple high-frequency detail information regions.

[0040] like Figure 2 As shown, this embodiment of the invention performs three-level wavelet decomposition on the acquired small target image. The first-level wavelet decomposition decomposes the small target image into four sub-images, namely the first-level low-frequency information region and three high-frequency detail information regions. In this embodiment of the invention, these three high-frequency detail information regions are referred to as the first-level high-frequency detail information region A, the first-level high-frequency detail information region B, and the first-level high-frequency detail information region C, respectively. The first-level high-frequency detail information regions A, B, and C are the horizontal high-frequency detail information region, the vertical high-frequency detail information region, and the diagonal high-frequency detail information region, respectively.

[0041] The second-level wavelet decomposition divides the first-level low-frequency information region into four sub-images: the second-level low-frequency information region and three high-frequency detail information regions. In this embodiment, the three high-frequency detail information regions obtained by the second-level wavelet decomposition are referred to as second-level high-frequency detail information region A, second-level high-frequency detail information region B, and second-level high-frequency detail information region C. Second-level high-frequency detail information regions A, B, and C are respectively a horizontal high-frequency detail information region, a vertical high-frequency detail information region, and a diagonal high-frequency detail information region.

[0042] The third level wavelet decomposition divides the second layer low frequency information region into four sub-images, which are third layer low frequency information region and three high frequency detail information regions. The three high frequency detail information regions obtained by the third level wavelet decomposition are referred to as third layer high frequency detail information region A, third layer high frequency detail information region B and third layer high frequency detail information region C. The third layer high frequency detail information region A, third layer high frequency detail information region B and third layer high frequency detail information region C are horizontal high frequency detail information region, vertical high frequency detail information region and diagonal high frequency detail information region respectively.

[0043] After the three level wavelet decomposition, the third layer low frequency information region and nine high frequency detail information regions are obtained. The nine high frequency detail information regions are first layer high frequency detail information region A, first layer high frequency detail information region B, first layer high frequency detail information region C, second layer high frequency detail information region A, second layer high frequency detail information region B, second layer high frequency detail information region C, third layer high frequency detail information region A, third layer high frequency detail information region B and third layer high frequency detail information region C.

[0044] In the three level wavelet decomposition, each layer low frequency information region contains the overall outline and general structure information of the small target image, and each layer high frequency detail information region contains more detailed information, such as edge information and texture information.

[0045] S2: performing image enhancement processing on the plurality of high frequency detail information regions, and performing image reconstruction on the image enhanced plurality of high frequency detail information regions and the third layer low frequency information region to obtain a reconstructed image.

[0046] As shown in Figure 2 , the nine high frequency detail information regions are subjected to image enhancement processing to improve the contrast between the unmanned aerial vehicle and the background in the small target image. The image enhancement is implemented by using gamma transformation. The specific process of image enhancement is as follows:

[0047] Four gray value thresholds are set, which are referred to as first threshold , second threshold , third threshold and fourth threshold . The first threshold , second threshold , third threshold and fourth threshold are used to process the nine high frequency detail information regions one by one as follows:

[0048] Taking the first layer high frequency detail information region A as an example, the following processing is performed: This represents the high-frequency detail information in the first-layer high-frequency detail information region A. To apply gamma transformation, it is typically necessary to extract the high-frequency detail information from region A before performing the gamma transformation. Normalization to Within the range, the high-frequency detail information region A of the first layer is then subjected to nonlinear compression or nonlinear stretching through gamma transformation: at the third threshold and the first threshold Within the interval, nonlinear compression is performed using gamma transformation to reduce the influence of non-contour information such as light and shadow, making non-edge regions smoother, at the first threshold. To the second threshold Third threshold To the fourth threshold Nonlinear stretching using gamma transformation within the interval makes the detailed areas clearer, especially when the value is greater than the second threshold. and less than the fourth threshold No processing is performed within the specified range to prevent excessive stretching that could lead to image distortion.

[0049] The standard expression for the gamma transformation is:

[0050] ,

[0051] in, This indicates the high-frequency detail information output. This represents high-frequency detail information from the input. Represents a proportionality constant, typically , This represents the gamma parameter. The value of determines the intensity and direction of the gamma transformation, controlling the degree of nonlinearity of the gamma transformation. This will compress the details of the first-layer high-frequency detail information region A. This will enhance the details of the first-layer high-frequency detail information region A. The gamma transformation expression in this embodiment of the invention is:

[0052] ,

[0053] in, This indicates the high-frequency detail information output. Indicates a certain value greater than 1° The fixed parameter values ​​are used to nonlinearly compress high-frequency detail information to suppress weak noise and subtle lighting changes. It represents a value greater than 0° and less than 1°. With fixed parameter values, high-frequency detail information is nonlinearly stretched to enhance edge details.

[0054] It should be noted that when , the first layer high frequency detail information area A has no change. All the original intensity and contrast of high frequency detail information are reserved; when , it belongs to an extreme nonlinear mapping, that is, all non-zero high frequency coefficients are raised to the maximum value, in practical applications, almost never use , but use a very small value close to 0 (for example or ) to achieve strong detail enhancement effect, while not completely losing intensity information. In addition, when (for example or ) can suppress weak noise and fine texture.

[0055] After completing the gamma transformation, the high frequency detail information is restored to the original gray value range, that is, 0-255. The high frequency detail information of the first layer high frequency detail information area A is enhanced after the gamma transformation, and the enhanced high frequency detail information is used to replace the original high frequency detail information in the first layer high frequency detail information area A, so as to realize the enhancement of the first layer high frequency detail information area A. Through the gamma transformation, the contrast between the high frequency detail information corresponding to the small target of the unmanned aerial vehicle and the background in the first layer high frequency detail information area A is enhanced.

[0056] The image enhancement process of the first layer high frequency detail information area A is performed on the other eight high frequency detail information areas one by one, and finally the enhancement of the nine high frequency detail information areas is realized. Since the third layer low frequency information area contains the main structure information of the small target image, the third layer low frequency information area is kept unchanged, and then the third layer low frequency information area and the nine enhanced high frequency detail information areas are reconstructed by inverse transformation to obtain the reconstructed image.

[0057] S3: improve the YOLOv5 model by using the CBAM attention module, and train the improved YOLOv5 model to obtain the improved YOLOv5 model.

[0058] The structure of the YOLOv5 model is shown in Figure 3 , including a backbone network, a neck network and a detection head, Figure 3 , wherein input represents input, conv represents convolution, concat represents splicing, conv2d represents two-dimensional convolution, upsample represents up-sampling, and SPPF represents fast pyramid pooling. The embodiments of the present application improve the YOLOv5 model, and then apply the improved YOLOv5 model to the reconstructed image recognition. The improvement method of the YOLOv5 model is as follows:

[0059] The CBAM attention module is inserted into the backbone of the YOLOv5 model, and the specific insertion position is the first C3 module after the input (input). The structure of the CBAM attention module is as shown in Figure 4 Since the small target of the unmanned aerial vehicle is usually more obvious in low-level features (such as edges and textures), the CBAM attention module is inserted into the shallow part of the YOLOv5 model, the attention of the YOLOv5 model to low-level features is enhanced through the CBAM attention module, refined features are obtained, and the detection effect of the small target of the unmanned aerial vehicle is improved. Moreover, the parameter amount and the calculation amount of the CBAM attention module are relatively small, and the training and inference speed of the YOLOv5 model is not greatly affected.

[0060] After the CBAM attention module is inserted into the YOLOv5 model, the improvement of the YOLOv5 model is completed, and the improved YOLOv5 model needs to be trained. The existing small target image is applied during the training. In the embodiment of the present application, the small target image of the unmanned aerial vehicle existing on the network or the small target image of the unmanned aerial vehicle in the public data set of the small target of the unmanned aerial vehicle is used as the training data, and the improved YOLOv5 model is trained. Before training, the small target image of the unmanned aerial vehicle in the training data is processed according to the method of step S1 to obtain the corresponding reconstruction image of the training data, and the improved YOLOv5 model is trained using the corresponding reconstruction image of the training data to obtain the YOLOv5 improved model.

[0061] For the improvement method of the YOLOv5 model, when the CBAM attention module is added to the backbone of the YOLOv5 model, other embodiments of the present application can add the CBAM attention module to the C3 module at other positions in the backbone, or add multiple CBAM attention modules to multiple C3 modules. Through reasonable debugging, the dual strategy of image enhancement combined with the CBAM attention module can also be realized.

[0062] S4: The small target of the reconstruction image is identified by using the YOLOv5 improved model.

[0063] The reconstruction image obtained in step S1 is input into the YOLOv5 improved model, and then the small target of the small target image is identified.

[0064] S5: The detection box position information corresponding to the continuous multiple frames of reconstruction images is obtained, the motion vector of the small target is calculated according to the detection box position information, and the acceleration module length is calculated according to the motion vector; whether the identified small target is a real target is judged according to the acceleration module length.

[0065] In the unmanned aerial vehicle small target detection, the small target usually occupies fewer pixels, and is easily affected by the background complexity, light change, shielding and other factors. When the small target is detected, noise or background and other elements are easily misjudged as the target, and the false detection phenomenon is generated. The embodiment of the application identifies and eliminates the false detection phenomenon, and improves the accuracy of unmanned aerial vehicle small target identification.

[0066] When the unmanned aerial vehicle small target identification is performed for a scene, usually dozens of frames of pictures of the scene are collected per second. For example, about 50 frames of small target images are collected per second in the embodiment of the application. If there is a real unmanned aerial vehicle small target in the identified small target image, the acceleration of the real unmanned aerial vehicle will not change dramatically in the adjacent frame small target image, and the corresponding acceleration module length will remain stable. According to the characteristics, whether the identified unmanned aerial vehicle small target is a real target is judged by judging the acceleration module length change of the identified small target.

[0067] The specific false detection judgment method is as follows:

[0068] The position information of the detection frame corresponding to the continuous multiple frames of reconstructed images is obtained (the reconstructed image has completed small target identification, and the detection frame position information exists), the target motion vector information including the velocity vector and the acceleration vector is obtained by using the difference calculation according to the detection frame position information, the acceleration module length of the small target corresponding to the target motion vector information is calculated according to the target motion vector information, and the false detection is judged according to the acceleration module length. The detection frame refers to the area of the unmanned aerial vehicle small target in the reconstructed image, the detection frame is selected as the boundary of the identified small target, the detection frame has position information, and the small target in the detection frame may be a real target or a non-real target. In theory, after a target appears in the small target image and is identified, the small target exists in each frame of the small target image, unless the small target disappears in the field of view. Correspondingly, the small target exists in each frame of the small target image. A detection frame corresponding to the position information exists in a series of small target images.

[0069] In the small target identification process of the five frames of reconstructed images by the YOLOv5 improved model, the YOLOv5 improved model generates a detection frame, and the position information of the detection frame of the five frames of reconstructed images is obtained. The detection frame position is represented by two-dimensional coordinates , represents the horizontal position of the detection frame in the reconstructed image, represents the vertical position of the detection frame in the reconstructed image. The position information of the detection frame in the five frames of reconstructed images is respectively represented by , , , and .

[0070] In any five-frame reconstructed image, three consecutive frames are selected to form a group of reconstructed images, which can form three groups, i.e., the first frame, the second frame and the third frame form a group, the second frame, the third frame and the fourth frame form a group, and the third frame, the fourth frame and the fifth frame form a group. The corresponding three detection box position information in each group of reconstructed images is obtained and difference calculation is performed, the velocity vector is calculated first, and then the acceleration vector is calculated according to the velocity vector. In the five-frame reconstructed image, the velocity vector is obtained by difference calculation of the position information of the adjacent two detection boxes:

[0071] ,

[0072] wherein the position information of the detection box in the reconstructed image is represented as , the sequence number of the frame is represented as , and the corresponding velocity vector is represented as . The three groups of acceleration vectors are The difference calculation formula is as follows:

[0073] ,

[0074] The acceleration module length is calculated according to the acceleration vector

[0075] ,

[0076] wherein, since the five-frame reconstructed image can only correspond to calculate three groups of acceleration vectors, the sequence number in the acceleration vector is only 1, 2 or 3. In the corresponding acceleration module length , the sequence number is only 1, 2 or 3.

[0077] The above operation is performed on each five-frame reconstructed image to obtain a series of acceleration module lengths, a stable threshold is set, and the sample variance of the acceleration module length is calculated:

[0078] ;

[0079] wherein, , and represent the first acceleration module length, the second acceleration module length and the third acceleration module length corresponding to the three groups of acceleration vectors, represents the average value of the acceleration module length, i.e. . ​

[0080] the sample variance of the acceleration module length is less than a set stable threshold value . If the sample variance of the acceleration module length is not greater than the stable threshold value , that is , it is determined that the corresponding small target of the unmanned aerial vehicle is a real target, and the small target framed by the corresponding detection frame does not exist false detection; if the sample variance of the acceleration module length is greater than the stable threshold value , that is , it is determined that the corresponding small target of the unmanned aerial vehicle is a non-real target, and the small target framed by the corresponding detection frame exists false detection.

[0081] When it is necessary to eliminate non-real targets, the detection frame corresponding to the small target determined as false detection can be eliminated.

[0082] The embodiment of the application carries out simulation experiment on the small target recognition method of the unmanned aerial vehicle based on wavelet decomposition and motion vector, and the experimental conditions are set to a shooting speed of fifty frames per second, the small target of the unmanned aerial vehicle is continuously shot, and theoretically the acceleration module length of the target remains unchanged in a short time, and there is a very large numerical difference with the variance caused by false detection. As shown in Figure 5 , it is a small target schematic diagram recognized by the improved YOLOv5 model, wherein uav represents an unmanned aerial vehicle, and the real target and the non-real target obtained after simulation and false detection judgment are as shown in Figure 6 , the horizontal coordinate (i.e. X coordinate) corresponds to the position index of the pixel in the horizontal direction (from left to right) of the image, and the vertical coordinate (i.e. Y coordinate) corresponds to the position index of the pixel in the vertical direction (from bottom to top) of the image, and the combination of the two can determine the two-dimensional space position of each pixel in the image. Figure 6 , the target correct position information is the detection frame position information corresponding to the real target, and the target false detection position information is the detection frame position information corresponding to the non-real target.

[0083] Although the embodiments of the application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the application. Those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the application.

[0084] The specific embodiments of the application described above do not constitute a limitation on the protection scope of the application. Any various other corresponding changes and modifications made according to the technical concept of the application should be included in the protection scope of the claims of the application.​​​

Claims

1. A small target recognition method for unmanned aerial vehicles based on wavelet decomposition and motion vectors, characterized by, The method comprises the following steps: S1: performing three-level wavelet decomposition on a small target image to obtain a third-level low-frequency information region and a plurality of high-frequency detail information regions; S2: performing image enhancement processing on the plurality of high-frequency detail information regions, and reconstructing the image by combining the plurality of high-frequency detail information regions after image enhancement with the third-level low-frequency information region to obtain a reconstructed image; S3: improving a YOLOv5 model by using a CBAM attention module and training the improved YOLOv5 model to obtain a YOLOv5 improved model; S4: identifying small targets in the reconstructed image by using the YOLOv5 improved model; S5: obtaining detection box position information corresponding to a plurality of continuous frames of reconstructed images, calculating a motion vector of the small target according to the detection box position information, calculating an acceleration module length according to the motion vector, and determining whether the identified small target is a real target according to the acceleration module length. 2.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The three-level wavelet decomposition process comprises the following steps: performing first-level wavelet decomposition on a small target image to obtain a first-level low-frequency information region; performing second-level wavelet decomposition on the first-level low-frequency information region to obtain a second-level low-frequency information region; and performing third-level wavelet decomposition on the second-level low-frequency information region to obtain a third-level low-frequency information region. 3.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The image enhancement is realized by gamma transformation. 4.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The method for improving the YOLOv5 model by using the CBAM attention module comprises the following steps: inserting the CBAM attention module into the backbone of the YOLOv5 model, and specifically inserting the CBAM attention module into the first C3 module after input. 5.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The training method of the improved YOLOv5 model comprises the following steps: obtaining an existing small target image, obtaining a reconstructed image corresponding to the existing small target image by the method of steps S1 and S2, and training the improved YOLOv5 model by using the reconstructed image corresponding to the existing small target image. 6.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The method for calculating the acceleration module length comprises the following steps: calculating a velocity vector according to adjacent detection box position information, calculating an acceleration vector according to adjacent velocity vectors, and calculating an acceleration module length according to the acceleration vector.

7. The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 6, characterized in that, The judging method of the real target is: setting a stability threshold If the sample variance of the acceleration module is not greater than the stability threshold , the corresponding target is a real target. 8.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 7, wherein, the sample variance of the acceleration modulus is: ； wherein , and respectively represent the first, second and third acceleration module lengths, represents the average value of the acceleration module lengths. 9.The UAV small target recognition method based on wavelet decomposition and motion vector according to claim 1, wherein, The number of the plurality of continuous frames of reconstructed images is five.

Images