A small sample target detection method and system for small unmanned aerial vehicle platform

Abstract

The present invention discloses a small sample target detection method and system applied to a small unmanned aerial vehicle platform. The small sample target detection method applied to a small unmanned aerial vehicle platform of the present invention comprises obtaining a sample image x to be identified collected by the small unmanned aerial vehicle platform. q And by generating candidate target regions, we get a set of candidate target region images, extract image features from any j-th candidate target region image and input the support vector machine model parameters Ω SVM The corresponding category detection results are obtained, and the detection results of N candidate target area images are output. Under the condition of limited computing resources, the present invention can realize online target detection based on a small number of labeled samples, and has the advantages of low dependence on hardware computing resources, and good effectiveness and real-time performance.

Description

Technical Field

[0001] The present invention belongs to target detection technology, and in particular relates to a small sample target detection method and system applied to a small unmanned aerial vehicle platform. Background Art

[0002] With the characteristics of flexibility, low cost, convenience and speed, small drones have shown irreplaceable air superiority in applications such as ground reconnaissance and aerial photography detection. Due to the limited payload capacity and energy supply of small drones, the onboard host computer is usually required to be light in weight and low in energy consumption, and cannot be equipped with high-energy-consuming devices such as conventional GPUs. Therefore, the computing power of the onboard platform is usually limited. At present, target detection technology based on deep learning usually adopts end-to-end deep and complex models, which not only has high requirements on the computing resources of the hardware platform, but also usually requires a large number of labeled samples for model pre-training for the target type to be detected, and then conducts online target detection. In actual applications, the target type is usually unknown, and it is difficult to provide a large number of labeled samples in advance. Therefore, small sample target detection technology with low computing resource dependence and the ability to perform online detection with a small number of labeled samples has become a bottleneck technology for the actual application of small drone ground target detection. Summary of the invention

[0003] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a small sample target detection method and system applied to a small unmanned aerial vehicle platform are provided. Under the condition of limited computing resources, the present invention can realize online target detection based on a small number of labeled samples, and has the advantages of low dependence on hardware computing resources, and good effectiveness and real-time performance.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A small sample target detection method applied to a small unmanned aerial vehicle platform, comprising:

[0006] 1) Obtain the sample image x to be identified collected by the small unmanned aerial vehicle platform q ;

[0007] 2) Treat the sample image x for identification q Generate candidate target regions and obtain a set of candidate target region images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI area coordinates, N is the number of candidate target area images;

[0008] 3) For any j-th candidate target area image Extracting image features A total of N image features are obtained

[0009] 4) N ​​image features Input the support vector machine model parameters Ω respectively SVM Get the corresponding category detection results Finally, the detection results of N candidate target area images are output

[0010] Optionally, in step 3), for any j-th candidate target area image Extracting image features It refers to the use of pre-trained convolutional neural network model Φ CNN Extract image features

[0011] Optionally, the convolutional neural network model Φ CNN It includes a feature extraction unit and a fully connected layer connected to each other, wherein the feature extraction unit includes a convolution layer, a downsampling pooling layer, a convolution layer and a downsampling pooling layer connected in sequence.

[0012] Optionally, step 2) treats the sample image x to be identified q Generating candidate target regions includes:

[0013] 2.1) The sample image x to be identified q Grayscale, saliency detection and binarization are performed in sequence to obtain the corresponding binary image;

[0014] 2.2) According to the sample image x to be identified q The corresponding binary image is used for contour detection and the sample images to be identified are extracted respectively. q Each candidate target area image And its corresponding ROI area coordinates Thus, we get a set of candidate target area images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI area coordinates, N is the number of candidate target area images.

[0015] Optionally, the saliency detection in step 2.1) includes: using a method based on global contrast, by calculating the global contrast of a single pixel on the entire image as the saliency value of the pixel, thereby obtaining a saliency grayscale image expressed by the saliency values ​​of each pixel.

[0016] Optionally, the calculation function expression of calculating the global contrast of a single pixel on the whole image as the saliency value of the pixel is:

[0017]

[0018] In the above formula, S(I k ) is the saliency value of the kth pixel in image I, I k is the gray value of the kth pixel in image I, I i is the gray value of the i-th pixel in image I, image I is the sample image x to be identified q .

[0019] Optionally, step 2.2) includes:

[0020] 2.2.1) For the sample image x to be identified q The corresponding binary image is used to obtain all contours in the binary image through contour detection, and the contours are sorted in descending order according to their lengths, and the top N contours with the largest contour lengths are selected as candidate target contours;

[0021] 2.2.2) Extract the minimum bounding rectangle of the N candidate target contours to form a preliminary candidate target region, and generate a candidate target detection frame by expanding the boundary of the preliminary candidate target region;

[0022] 2.2.3) Treat the sample image x as the identification q There are multiple overlapping candidate target detection frames in the space to merge and generate the final candidate target area, so as to determine the ROI area coordinates of each candidate target area And by q Each candidate target area image is cropped from the original image Thus, we get a set of candidate target area images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI area coordinates, N is the number of candidate target area images.

[0023] Optionally, before step 4), it also includes constructing and training a support vector machine classification model Ω SVM Steps and training support vector machine classification model Ω SVM include:

[0024] S1) Based on the input prior sample x containing the target type to be identified s , label annotation and obtain M label samples in is the prior sample x s The i-th target area image in, is the i-th target area image The category label of any i-th target area image Extracting image features A total of M image features and their label pairs are obtained

[0025] S2) Based on M image features and their label pairs Train the support vector machine model parameters Ω SVM , thereby establishing a mapping relationship between image features and categories, which can be used to realize image target detection of small unmanned aerial vehicle platforms.

[0026] In addition, the present invention also provides a small sample target detection system applied to a small unmanned aerial vehicle platform, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform.

[0027] In addition, the present invention also provides a computer-readable storage medium, in which a computer program is stored, and the computer program is used to be programmed or configured by a microprocessor to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform.

[0028] Compared with the prior art, the present invention has the following advantages: the small sample target detection method applied to a small unmanned aerial vehicle platform of the present invention comprises obtaining a sample image x to be identified acquired by the small unmanned aerial vehicle platform q And by generating candidate target regions, we get a set of candidate target region images, extract image features from any j-th candidate target region image and input the support vector machine model parameters Ω SVM The corresponding category detection results are obtained, and the detection results of N candidate target area images are output. Under the condition of limited computing resources, the present invention does not rely on high-energy-consuming devices such as conventional GPUs, can realize online target detection based on a small number of labeled samples, and has the advantages of low dependence on hardware computing resources, taking into account both target detection performance and computing real-time, and having good effectiveness and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the basic flow of the method of the embodiment of the present invention.

[0030] Figure 2 is the convolutional neural network model Φ in the embodiment of the present invention. CNN Schematic diagram of the network structure.

[0031] Figure 3This is a schematic diagram of the principle of generating candidate target areas in an embodiment of the present invention.

[0032] Figure 4 It is a complete schematic diagram of the application and training principles of the method of the embodiment of the present invention.

[0033] Figure 5 This is a typical example of a small UAV ground target reconnaissance detection dataset in an embodiment of the present invention.

[0034] Figure 6 Comparison of recall rate statistics of small sample target detection in the embodiments of the present invention.

[0035] Figure 7 The following is a comparison of the accuracy statistics of small sample target detection in the embodiments of the present invention. DETAILED DESCRIPTION

[0036] like Figure 1 The small sample target detection method applied to a small unmanned aerial vehicle platform in this embodiment includes:

[0037] 1) Obtain the sample image x to be identified collected by the small unmanned aerial vehicle platform q ;

[0038] 2) Treat the sample image x for identification q Generate candidate target regions and obtain a set of candidate target region images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI (region of interest) coordinates, N is the number of candidate target region images;

[0039] 3) For any j-th candidate target area image Extracting image features A total of N image features are obtained

[0040] 4) N ​​image features Input the support vector machine model parameters Ω respectively SVM Get the corresponding category detection results Finally, the detection results of N candidate target area images are output

[0041] In this embodiment, in step 3), for any j-th candidate target area image Extracting image features It refers to the use of pre-trained convolutional neural network model Φ CNN Extract image features The method of this embodiment is aimed at practical applications such as small UAV ground target reconnaissance and detection. The candidate target area in the real-time data stream is obtained by saliency detection, and the image features of the candidate target area are extracted by using a pre-trained convolutional neural network. Finally, a lightweight SVM classifier (support vector machine model parameter Ω SVM ) can realize online detection of unknown types of targets under the condition of a small number of labeled samples. At the same time, the algorithm has little dependence on hardware computing resources and has both effectiveness and real-time performance.

[0042] In view of the successful application of convolutional neural networks in the current field of computer vision, this embodiment uses a convolutional neural network (CNN) model pre-trained on a large-scale data set to extract candidate target areas and support sample image feature extraction. The convolutional neural network model used to extract image features can use a feasible network structure as needed. As an optional implementation method, based on the characteristics of small unmanned aerial vehicle platforms, such as Figure 2 As shown, the convolutional neural network model Φ used in this embodiment CNN It includes interconnected feature extraction units and fully connected layers. The feature extraction unit includes sequentially connected convolutional layers, downsampling pooling layers, convolutional layers, and downsampling pooling layers. The pre-trained convolutional neural network model extracts features from image data through local convolution, downsampling pooling, and full connection operations to form a high-dimensional vector representation, which has good image feature representation capabilities. Convolutional neural network model Φ CNN The function form can be expressed as: v = Φ CNN (x), where for any region image x, a convolutional neural network Φ is used CNN Extracted high-dimensional feature vectors. Through regional feature extraction, a small number of support samples and candidate target region images are encoded into fixed-dimensional sample representation vectors. Afterwards, a support vector machine (SVM) classification model is constructed, and the classifier model parameters are trained based on the feature vectors of a small number of support samples, and then the candidate target feature vectors are classified and recognized online.

[0043] See also Figure 3 , step 2) treat the identified sample image x q Generating candidate target regions includes:

[0044] 2.1) The sample image x to be identified q Graying, saliency detection and binarization are performed in sequence to obtain the corresponding binary image; wherein, graying is to convert the original color image into a single-channel gray image, saliency detection is used to enhance the image of the candidate target area, and after the saliency gray image is obtained by saliency detection, it is binarized by setting a threshold to form a binary image;

[0045] 2.2) According to the sample image x to be identified q The corresponding binary image is used for contour detection and the sample images to be identified are extracted respectively. q Each candidate target area image And its corresponding ROI area coordinates Thus, we get a set of candidate target area images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI area coordinates, N is the number of candidate target area images.

[0046] For the real-time image data acquired by the drone, it is first necessary to detect the candidate target areas that may exist therein. This embodiment adopts a candidate target area generation method based on saliency detection. The saliency detection in step 2.1) includes: using a method based on global contrast, by calculating the global contrast of a single pixel on the whole image as the saliency value of the pixel, thereby obtaining a saliency grayscale image expressed by the saliency value of each pixel. Specifically, the calculation function expression of calculating the global contrast of a single pixel on the whole image as the saliency value of the pixel is:

[0047]

[0048] In the above formula, S(I k ) is the saliency value of the kth pixel in image I, I k is the gray value of the kth pixel in image I, I i is the gray value of the i-th pixel in image I, image I is the sample image x to be identified q .

[0049] See also Figure 3 , step 2.2) comprises:

[0050] 2.2.1) For the sample image x to be identified q The corresponding binary image is detected by contour detection to obtain all contours in the binary image, and the contours are sorted in descending order according to their lengths, and the top N (TopN) contours with the largest contour lengths are selected as candidate target contours;

[0051] 2.2.2) Extract the minimum bounding rectangle of the N candidate target contours to form a preliminary candidate target region, and generate a candidate target detection frame by padding the preliminary candidate target region. The size of the padding can be specified based on experience. Since the minimum bounding rectangle of the extracted contour is usually difficult to completely contain the target, the candidate target detection frame is generated by padding based on the preliminary candidate target region.

[0052] 2.2.3) Treat the sample image x as the identification q There are multiple overlapping candidate target detection frames in the space to merge and generate the final candidate target area, so as to determine the ROI area coordinates of each candidate target area And by q Each candidate target area image is cropped from the original image Thus, we get a set of candidate target area images in is the sample image x to be identified q The j-th candidate target region image, for The corresponding ROI area coordinates, N is the number of candidate target area images.

[0053] like Figure 4 As shown, for small sample target detection, in the real-time target detection process, the user needs to provide a small number of support samples, that is, a small number of label samples are annotated online for the target to be identified, for subsequent target classification and identification. Specifically, before step 4) of this embodiment, it also includes constructing and training a support vector machine classification model Ω SVM Steps and training support vector machine classification model Ω SVM include:

[0054] S1) Based on the input prior sample x containing the target type to be identified s , label annotation and obtain M label samples in is the prior sample x s The i-th target area image in, is the i-th target area image The category label of any i-th target area image Extracting image features A total of M image features and their label pairs are obtained

[0055] S2) Based on M image features and their label pairs Train the support vector machine model parameters Ω SVM, thereby establishing a mapping relationship between image features and categories, which is used to realize image target detection of small unmanned aerial vehicle platforms. Similar to conventional small sample learning, the small sample target detection process includes support samples (i.e., the prior samples x in the previous article) s ) and test samples (sample images to be identified x q ), that is, for the target type to be identified, a small number of label samples are annotated online based on prior knowledge, and on this basis, the target in the real-time data stream is detected online. Using the trained support vector machine model Ω SVM For candidate target area For classification and recognition, it can be expressed as: Combined with candidate target area coordinates The final output online test results can be obtained

[0056] In order to verify the effectiveness of the algorithm and system, in this embodiment, an actual flight data set is collected for the typical application of small UAV ground target reconnaissance. Among them, the ground targets use highly realistic moving military models, including tanks, missile launchers, command vehicles, light trucks, fighters, transport aircraft, helicopters, early warning aircraft, destroyers, combat boats and other targets. Small UAVs are used for ground aerial reconnaissance to obtain a large amount of moving target data, such as Figure 5 In particular, this embodiment targets the typical scenario where the computing resources of the small drone-borne platform are limited. The embedded computing platform RK3399Pro, which can be directly deployed on various small drones, is selected as the onboard host computer to carry out algorithm and system case analysis. The CPU computing resources of the embedded computing platform RK3399Pro are dual-core A72 processors and quad-core A53 processors, and its dynamic running memory is only 200K, which is difficult to support the online operation of conventional large-scale complex deep learning models.

[0057] Based on the above-mentioned UAV ground reconnaissance real data and airborne embedded computing platform, a performance test experiment was carried out for the method proposed in this embodiment. The experiment uses the commonly used accuracy (Precise) and recall (Recall) in target detection as performance evaluation indicators, and counts the number of detected targets under a certain intersection over Union (IoU) threshold, where the accuracy represents the proportion of correctly detected targets to the total number of detected targets, and the recall represents the proportion of correctly detected targets to the total number of true targets. In the regional feature extraction stage, this example uses the Alexnet convolutional neural network pre-trained on the ImageNet dataset to extract candidate target regions and support sample image features, and uses the output of the 7th fully connected layer (FC7) of the pre-trained model as the regional image feature. The experiment uses the M-way N-shot experimental setting commonly used in small sample learning, and conducts performance testing in constructing multiple training-test batch data iterations. In order to ensure the reliability of the experimental results, a large number of random experimental statistical results are adopted. That is, 600 groups of batch data are randomly selected in each experiment (each group of batch data contains 5 / 10-way 5 / 10-shot 10-query samples). The final experimental results are obtained by statistics on 600 groups of random batch data, including the average recognition accuracy, recall rate and its 95% confidence interval, as shown in Table 1.

[0058] Table 1: Statistics of small sample target detection results under different experimental conditions.

[0059]

[0060] In Table 1, TopN is the number of candidate target regions generated by saliency detection. The intersection-over-union ratio (IOU) of positive and negative sample judgment in the target detection process is set to 0.3. In order to more intuitively compare the target detection results under various parameter settings, the results listed in Table 1 are plotted as follows: Figure 6 and Figure 7 As shown, Figure 6 is the comparison curve of target detection recall rate, Figure 7 This is the comparison curve of target detection accuracy. Figure 6 and Figure 7It can be seen that as the number of candidate target regions TopN increases during saliency detection, the target detection recall rate gradually increases, while the target detection accuracy rate decreases, which is consistent with the performance of conventional supervised learning target detection methods; at the same time, for small sample learning, the target detection accuracy and recall rate both increase with the increase of sample number, and decrease with the increase of classification categories, which is consistent with the performance of small sample target classification. It is worth mentioning that the method proposed in this embodiment achieved a recall rate of up to 62.4% (5-way 10-shot, TopN=8) and a detection accuracy rate of 27.3% (5-way 10-shot, TopN=4) in the small sample target detection experiment, indicating that the small sample target detection method proposed in this embodiment can achieve effective detection and recognition of new categories of targets under the support of a small number of labeled samples.

[0061] In addition to the above-mentioned small sample target detection accuracy evaluation, this embodiment also carried out an airborne embedded platform computing efficiency analysis, including the computational time efficiency of each step of the algorithm, as shown in Table 2.

[0062] Table 2: Computational efficiency and real-time analysis of the RK3399Pro airborne embedded computing platform.

[0063] Image resolution Candidate target region generation Regional feature extraction SVM feature classification 1280*720 ～1.92s ~3.41s / ROI ～0.001s

[0064] On the RK3399Pro processor with severely limited computing resources, the method of this embodiment can effectively carry out online target detection on high-definition images with a resolution of 1280*720, where the candidate target area generation takes about 1.92 seconds. Since the convolutional neural network model is used to calculate the image features during the regional feature extraction process, the calculation process is relatively complex and takes about 3.41s / ROI, while the SVM feature classification process only takes about 0.001 seconds. Overall, this algorithm can carry out online target detection on an airborne embedded platform and meet certain real-time requirements. It is worth mentioning that the small sample target detection method proposed in this embodiment is a method based on saliency detection and feature engineering, which belongs to a parameter-free model. Unlike the conventional method that contains a large number of parameters to be trained and needs to train parameters on a large number of base class label samples, this method does not require a large number of base class label sample training. It can realize online detection and recognition of new types of targets under small sample conditions based on pre-trained image feature extraction and parameter-free saliency detection. When users actually deploy the system in the later stage, there is no need to collect a large amount of labeled data for retraining of model parameters. Only a better image feature extractor in the field is needed for rapid deployment and online application. It also supports users to customize new category support samples online, and has small sample detection and sustainable learning capabilities for new category targets.

[0065] In summary, in view of the limitation of computing resources of small unmanned aerial vehicle platforms, the method of this embodiment combines deep learning feature engineering with shallow feature classifier methods, and proposes a small sample target detection method that takes into account both practical performance and computing efficiency, including the main steps of generating candidate target regions in reconnaissance images, extracting regional features, and classifying target features: first, the region of interest (ROI) in the real-time reconnaissance image is obtained through saliency detection, that is, the possible target region; then, the ROI regional features are extracted through a pre-trained deep learning model; finally, a support vector machine (SVM) model is constructed based on a small number of support sample features to perform feature classification and recognition on candidate targets in real-time data streams. In particular, this embodiment collects and organizes actual flight data sets for unmanned aerial vehicle ground reconnaissance applications, and conducts case analysis based on the embedded airborne processing platform RK3399Pro. The system can realize online target detection under limited computing resources, and has good effectiveness and real-time performance.

[0066] In addition, the present embodiment also provides a small sample target detection system applied to a small unmanned aerial vehicle platform, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform as described above. As described above, since the method of the present embodiment does not rely on high-energy-consuming devices such as conventional GPUs under the condition of limited computing resources, it can realize online target detection based on a small number of labeled samples, and has the advantages of low dependence on hardware computing resources, taking into account target detection performance and computing real-time, and having good effectiveness and real-time performance, the small sample target detection system applied to a small unmanned aerial vehicle platform in the above text can be a drone platform; however, it should be noted that even if it is implemented on a computer connected to the drone, it can also be realized under the condition of limited computing resources, without relying on high-energy-consuming devices such as conventional GPUs, and can realize online target detection based on a small number of labeled samples, and has low dependence on hardware computing resources, taking into account target detection performance and computing real-time performance, and having good effectiveness and real-time performance.

[0067] In addition, this embodiment also provides a computer-readable storage medium, in which a computer program is stored. The computer program is used to be programmed or configured by a microprocessor to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform as described above.

[0068] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0069] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A small sample target detection method applied to a small unmanned aerial vehicle platform, characterized in that include: 1) Obtain the sample image to be identified collected by the small UAV platform x q ; 2) Sample images to be identified x q Generate candidate target regions and obtain a set of candidate target region images {( x j q , R j q ), j =1,2,…, N }, in x j q is the sample image to be identified x q No. j candidate target region images, R j q for x j q The corresponding ROI area coordinates, N is the number of candidate target area images; 3) For any j Candidate target region images x j q Extracting image features v j s , a total of N Image features { v j s , i =1,2,…, N }; 4) N Image features { v j s , i =1,2,…, N } Input the support vector machine model parameters respectively Ω SVM Get the corresponding category detection results y j q , the final output N The detection results of candidate target area images {( y j q , R j q ), i =1,2,…, N }; Step 2) Sample image to be identified x q Generating candidate target regions includes: 2.1) The sample image to be identified x q Grayscale, saliency detection and binarization are performed in sequence to obtain the corresponding binary image; 2.2) According to the sample image to be identified x q The corresponding binary image is used for contour detection and the sample images to be identified are extracted respectively. x q Each candidate target area image x j q And its corresponding ROI area coordinates R j q , thus obtaining the candidate target area image set {( x j q , R j q ), j =1,2,…, N }, in x j q is the sample image to be identified x q No. j candidate target region images, R j q for x j q The corresponding ROI area coordinates, N is the number of candidate target area images; Step 2.2) includes: 2.2.1) For the sample image to be identified x q The corresponding binary image is obtained by contour detection to obtain all contours in the binary image, sort them in descending order according to their contour length, and select the ones with the largest contour length. N contours as candidate target contours; 2.2.2) Yes N The minimum circumscribed rectangular frame is extracted from the candidate target contours to form a preliminary candidate target area, and the candidate target detection frame is generated by expanding the boundary of the preliminary candidate target area; 2.2.3) Sample images to be identified x q There are multiple overlapping candidate target detection frames in the space to merge and generate the final candidate target area, so as to determine the ROI area coordinates of each candidate target area R j q , and by x q Each candidate target area image is cropped from the original image x j q , thus obtaining the candidate target area image set {( x j q , R j q ), j =1,2,…, N }, in x j q is the sample image to be identified x q No. j candidate target region images, R j q for x j q The corresponding ROI area coordinates, N is the number of candidate target region images.

2. The small sample target detection method applied to a small unmanned aerial vehicle platform according to claim 1 is characterized in that: In step 3), for any j Candidate target region images x j q Extracting image features v j s Refers to the use of pre-trained convolutional neural network model Φ CNN Extract image features v j s .

3. The small sample target detection method applied to a small unmanned aerial vehicle platform according to claim 2 is characterized in that: The convolutional neural network model Φ CNN It includes a feature extraction unit and a fully connected layer connected to each other, wherein the feature extraction unit includes a convolution layer, a downsampling pooling layer, a convolution layer and a downsampling pooling layer connected in sequence.

4. The small sample target detection method applied to a small unmanned aerial vehicle platform according to claim 1 is characterized in that: The saliency detection in step 2.1) includes: using a global contrast-based method to calculate the global contrast of a single pixel on the entire image as the saliency value of the pixel, thereby obtaining a saliency grayscale image expressed by the saliency values ​​of each pixel.

5. The small sample target detection method applied to a small unmanned aerial vehicle platform according to claim 4 is characterized in that: The calculation function expression of calculating the global contrast of a single pixel on the whole image as the saliency value of the pixel is: In the above formula, S ( I k ) is the image I The k The saliency value of pixels, I k For images I The k The gray value of a pixel, I i For images I The i The gray value of pixels, image I is the sample image to be identified x q .

6. The small sample target detection method applied to a small unmanned aerial vehicle platform according to claim 1 is characterized in that: Step 4) also includes constructing and training the support vector machine classification model Ω SVM Steps and training support vector machine classification model Ω SVM include: S1) Based on the input prior samples containing the target type to be identified x s , label and get M Label samples {( x i s , y i s ), i =1,2,…, M }, in x i s The prior sample x s The i target area image, y i s For the i Target area image x i s , respectively, for any i Target area image x i s Extracting image features v i s , a total of M image features and their label pairs {( v i s , y i s ), i =1,2,…, M }; S2) Based on M image features and their label pairs {( v i s , y i s ), i =1,2,…, M }Train the support vector machine model parameters Ω SVM , thereby establishing a mapping relationship between image features and categories, which can be used to realize image target detection of small unmanned aerial vehicle platforms.

7. A small sample target detection system applied to a small unmanned aerial vehicle platform, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored therein, characterized in that: The computer program is used to be programmed or configured by a microprocessor to execute the steps of the small sample target detection method applied to a small unmanned aerial vehicle platform as described in any one of claims 1 to 6.

Images