Small target recognition method, system, terminal and storage medium

Abstract

The present invention relates to the field of image recognition technology. The present invention discloses a small target recognition method, system, terminal and storage medium, which can improve the accuracy of small target recognition when the data domain of unlabeled images is uncertain. The small target recognition method includes acquiring multiple unlabeled images; using a first target detection module to perform small target recognition on each of the unlabeled images, and outputting a first recognition result, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and more specifically, to a small target recognition method, system, terminal and storage medium. Background Art

[0002] Traditional small target recognition methods usually use unsupervised domain adaptation (UDA) algorithms and semi-supervised domain generalization (SSDG) algorithms to obtain small target recognition results. Unsupervised domain adaptation algorithms and semi-supervised domain generalization algorithms usually rely on a clear understanding of the source data domain and the target data domain. Specifically, if the unlabeled image comes from the source data domain, the small target recognition task can be regarded as a semi-supervised problem, and the semi-supervised domain generalization algorithm is used to recognize small targets; if the unlabeled image belongs to the target data domain, the small target recognition task is regarded as an unsupervised problem, and the unsupervised domain adaptation algorithm is used to recognize small targets. However, in most application scenarios, the data domain of unlabeled images is uncertain, so it is difficult to select a suitable algorithm for small target recognition tasks, resulting in poor recognition results. Summary of the invention

[0003] The purpose of the embodiments of the present application is to provide a small target recognition method, system, terminal and storage medium, which can improve the accuracy of small target recognition when the data domain of unlabeled images is uncertain. The embodiments of the present application are mainly implemented through the following technical solutions:

[0004] A first aspect of an embodiment of the present application provides a small target recognition method, comprising:

[0005] Acquire multiple unlabeled images;

[0006] A first target detection module is used to perform small target recognition on each of the unlabeled images, and a first recognition result is output, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image.

[0007] According to an embodiment of the present application, the step of using the first target detection module to perform small target recognition on each of the unmarked images and outputting a first recognition result includes:

[0008] Using the first convolutional layer of the first target detection module to process the target unlabeled image and output a first feature map;

[0009] Processing the first feature map using the first convolutional layer regularization layer of the first target detection module to output a second feature map;

[0010] Using the first cross-stage module of the first object detection module to process the second feature map, and output a third feature map;

[0011] Processing the third feature map using the second convolutional layer regularization layer of the first target detection module to output a fourth feature map;

[0012] Processing the fourth feature map using the second cross-stage module of the first object detection module to output a fifth feature map;

[0013] Using the third convolutional layer regularization layer of the first target detection module to process the fifth feature map, and output a sixth feature map;

[0014] Using the third cross-stage module of the first object detection module to process the sixth feature map, and output a seventh feature map;

[0015] Processing the seventh feature map using the fourth convolutional layer regularization layer of the first target detection module to output an eighth feature map;

[0016] Processing the eighth feature map using the first pooling layer of the first object detection module to output a ninth feature map;

[0017] Processing the ninth feature map using a fourth cross-stage module of the first object detection module to output a tenth feature map;

[0018] Using the fifth convolutional layer regularization layer of the first target detection module to process the tenth feature map, and output an eleventh feature map;

[0019] Processing the eleventh feature map using the first upsampling layer of the first object detection module to output a twelfth feature map;

[0020] Using a fifth cross-stage module of the first object detection module to process the twelfth feature map, and output a thirteenth feature map;

[0021] Using the sixth convolutional layer regularization layer of the first target detection module to process the thirteenth feature map, and output a fourteenth feature map;

[0022] Processing the fourteenth feature map using the second upsampling layer of the first object detection module to output a fifteenth feature map;

[0023] Processing the fifteenth feature map using the second convolutional layer of the first target detection module, and outputting the first recognition result;

[0024] The target unlabeled image is any unlabeled image among the multiple unlabeled images.

[0025] According to an embodiment of the present application, the training method of the first target detection module includes:

[0026] Acquire a first training data set and the first training pseudo-label set, wherein each first training data in the first training data set has a one-to-one correspondence with one of the first training pseudo-labels in the first training pseudo-label set;

[0027] Inputting target first training data into a first initial detection module to obtain a first training result, wherein the target first training data is any first training data in the first training data set;

[0028] Performing a non-maximum suppression operation on the first training result to obtain a third pseudo label;

[0029] Calculate a first loss function based on a first training pseudo label corresponding to the target first training data and the third pseudo label;

[0030] adjusting parameters of the first initial detection module based on the first loss function to form the first target detection module;

[0031] All of the third pseudo labels constitute a second training pseudo label set.

[0032] According to one embodiment of the present application, the small target recognition method further includes:

[0033] Obtain multiple labeled images;

[0034] Using a second target detection module to perform small target recognition on each of the marked images, and output a second recognition result;

[0035] A non-maximum suppression operation is performed on the second recognition result to obtain a first pseudo label, and the first pseudo label is added to a first training pseudo label set, wherein the first training pseudo label set is used to train the first target detection module.

[0036] According to an embodiment of the present application, the step of using the second target detection module to perform small target recognition on each of the marked images and outputting a second recognition result includes:

[0037] Processing the target labeled image using the third convolutional layer of the second target detection module to output a sixteenth feature map;

[0038] Processing the sixteenth feature map using the seventh convolutional layer regularization layer of the second object detection module, and outputting a seventeenth feature map;

[0039] Using the sixth cross-stage module of the second object detection module to process the seventeenth feature map, and output an eighteenth feature map;

[0040] Using the eighth convolutional layer regularization layer of the second object detection module to process the eighteenth feature map, and output a nineteenth feature map;

[0041] Using the seventh cross-stage module of the second object detection module to process the nineteenth feature map, and output a twentieth feature map;

[0042] Using the ninth convolutional layer regularization layer of the second object detection module to process the 20th feature map, and output a 21st feature map;

[0043] Using the eighth cross-stage module of the second object detection module to process the twenty-first feature map, and output a twenty-second feature map;

[0044] Using the tenth convolutional layer regularization layer of the second object detection module to process the twenty-second feature map, and output a twenty-third feature map;

[0045] Processing the twenty-third feature map using the second pooling layer of the second object detection module, and outputting a twenty-fourth feature map;

[0046] Using a ninth cross-stage module of the second object detection module to process the twenty-fourth feature map, and output a twenty-fifth feature map;

[0047] Using the eleventh convolutional layer regularization layer of the second object detection module to process the twenty-fifth feature map, and output a twenty-sixth feature map;

[0048] Processing the twenty-sixth feature map using the third upsampling layer of the second object detection module, and outputting a twenty-seventh feature map;

[0049] Processing the twenty-seventh feature map using a tenth cross-stage module of the second object detection module to output a twenty-eighth feature map;

[0050] Processing the twenty-eighth feature map using the twelfth convolutional layer regularization layer of the second object detection module, and outputting a twenty-ninth feature map;

[0051] Processing the twenty-ninth feature map using the fourth upsampling layer of the second object detection module, and outputting a thirtieth feature map;

[0052] Processing the thirtieth feature map using the fourth convolutional layer of the second object detection module, and outputting the second recognition result;

[0053] The target labeled image is any labeled image among the multiple labeled images.

[0054] According to one embodiment of the present application, the training method of the second object detection module includes:

[0055] Acquire a second training data set and the second training pseudo-label set, wherein each second training data in the second training data set has a one-to-one correspondence with one of the second training pseudo-labels in the second training pseudo-label set;

[0056] Inputting the target second training data into the second initial detection module to obtain a second training result, wherein the target second training data is any second training data in the second training data set;

[0057] Performing a non-maximum suppression operation on the second training result to obtain a fourth pseudo label;

[0058] Calculate a second loss function based on a second training pseudo label corresponding to the target second training data and the fourth pseudo label;

[0059] adjusting parameters of the second initial detection module based on the second loss function to form the second target detection module;

[0060] All the fourth pseudo labels are added to the first training pseudo label set.

[0061] According to an embodiment of the present application, the training method of the first target detection module further includes:

[0062] Generate a first cross pseudo-supervision loss function based on the first loss function and the second loss function;

[0063] Adjust parameters of the first object detection module based on the first cross pseudo-supervision loss function.

[0064] According to an embodiment of the present application, the training method of the second object detection module further includes:

[0065] Generate a second cross pseudo-supervision loss function based on the first loss function and the second loss function;

[0066] Adjust parameters of the second object detection module based on the second cross pseudo-supervision loss function.

[0067] A second aspect of an embodiment of the present application provides a small target recognition system, including:

[0068] An acquisition module, used for acquiring a plurality of unlabeled images;

[0069] The first recognition result output module is used to use the first target detection module to perform small target recognition on each of the unlabeled images and output a first recognition result, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image.

[0070] The third aspect of an embodiment of the present application provides a terminal device, including: a processor and a memory, the memory is used to store a computer program, the processor is used to call and run the computer program stored in the memory, and execute the steps of the small target recognition method provided in the first aspect of the embodiment of the present application.

[0071] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, which is used to store a computer program, and the computer program enables a computer to execute the steps of the small target recognition method provided in the first aspect of the embodiment of the present application.

[0072] The beneficial effects of the embodiments of the present application include at least:

[0073] The small target recognition method provided in the embodiment of the present application designs a first target detection module, and the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image, so that each unlabeled image can obtain a clear first recognition result after being recognized by the first target detection module. Therefore, compared with the prior art with poor recognition effect, the present application can improve the accuracy of small target recognition without determining the data domain of the unlabeled image. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0075] Figure 1 A comparison chart of generalization methods for different data domains;

[0076] Figure 2 A flowchart of a small target recognition method in some embodiments of the present invention;

[0077] Figure 3 is a block diagram of the architecture of the first target detection module and the second target detection module in some embodiments of the present invention;

[0078] Figure 4 A block diagram of the first convolutional layer regularization layer of the present invention in some embodiments;

[0079] Figure 5 is a block diagram of the first cross-stage module of the present invention in some embodiments;

[0080] Figure 6 It is a block diagram of a small target recognition system of the present invention in some embodiments. DETAILED DESCRIPTION

[0081] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0082] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0083] The terms "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0084] The terms "comprises," "includes," or any other variations thereof are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or elements is not necessarily limited to those steps or elements explicitly listed but may include other steps or elements not explicitly listed or inherent to such process, method, product, or apparatus.

[0085] Unless otherwise defined, all technical and scientific terms used in the specification of this application have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The term "and / or" used in the specification of this application includes any and all combinations of one or more related listed items.

[0086] Traditional small target recognition methods can be roughly divided into two categories, one of which is a recognition method based on segmentation algorithm, and the other is a recognition method based on object detection. The recognition method based on segmentation algorithm requires pixel-level labels, that is, each pixel must be correctly labeled as whether it is part of a small target. This form of labeling requires manual labeling of a large number of images, which will result in expensive labor costs. On the other hand, the recognition method based on object detection mainly focuses on identifying whether there are small targets in the image, and does not require pixel-level labeling. This type of method automatically identifies the presence of small targets by training object detectors and provides their corresponding positions and bounding boxes. Compared with the method based on segmentation algorithm, the recognition method based on object detection can reduce the labeling cost. Therefore, the inventor conducted research and development based on the recognition method based on object detection.

[0087] Traditional object detection-based recognition methods treat the small object recognition task as a supervised learning problem and fail to fully utilize the large amount of available unlabeled images (i.e., unlabeled source (training) data) to enhance cross-domain robustness. This shortcoming makes the trained object detector perform poorly when applied to target samples that show distribution drift. In many real-world scenarios, the data domain of unlabeled images is uncertain. Existing unsupervised domain adaptation algorithms and semi-supervised domain generalization algorithms are difficult to solve this problem. For details, please refer to Figure 1 As shown. Figure 1 In the case of unlabeled source (training) data, there is data domain uncertainty (i.e. Figure 1 The 4th row, 3rd column, and white data domain in the image makes it difficult to determine whether the real scene is a semi-supervised domain generalization problem or an unsupervised domain adaptation problem. Therefore, it is difficult to directly use the unsupervised domain adaptation algorithm or the semi-supervised domain generalization algorithm to perform small target recognition tasks, resulting in poor recognition results. In this regard, the embodiment of the present application proposes a small target recognition method to solve the above problem. The specific implementation methods of the present application are further described below in conjunction with the accompanying drawings.

[0088] like Figure 2 FIG. 1 is a flowchart of a small target recognition method provided by the first aspect of the embodiment of the present application. Figure 2 In the method, the small target recognition method comprises:

[0089] S1. Acquire multiple unlabeled images.

[0090] S2. Use a first target detection module to perform small target recognition on each of the unlabeled images and output a first recognition result, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image.

[0091] like Figure 3As shown, the first target detection module includes a first convolutional layer, a first convolutional layer regularization layer, a first cross-stage module, a second convolutional layer regularization layer, a second cross-stage module, a third convolutional layer regularization layer, a third cross-stage module, a fourth convolutional layer regularization layer, a first pooling layer, a fourth cross-stage module, a fifth convolutional layer regularization layer, a first upsampling layer, a fifth cross-stage module, a sixth convolutional layer regularization layer, a second upsampling layer and a second convolutional layer.

[0092] Further, the step of using the first target detection module to perform small target recognition on each of the unlabeled images and outputting a first recognition result includes: using the first convolutional layer to process the target unlabeled image and outputting a first feature map; using the first convolutional layer regularization layer to process the first feature map and outputting a second feature map; using the first cross-stage module to process the second feature map and outputting a third feature map; using the second convolutional layer regularization layer to process the third feature map and outputting a fourth feature map; using the second cross-stage module to process the fourth feature map and outputting a fifth feature map; using the third convolutional layer regularization layer to process the fifth feature map and outputting a sixth feature map; using the third cross-stage module to process the sixth feature map and outputting a seventh feature map; using the fourth convolutional layer regularization layer to process the seventh feature map and output The eighth feature map; the eighth feature map is processed by the first pooling layer to output a ninth feature map; the ninth feature map is processed by the fourth cross-stage module to output a tenth feature map; the tenth feature map is processed by the fifth convolutional layer regularization layer to output an eleventh feature map; the eleventh feature map is processed by the first upsampling layer to output a twelfth feature map; the twelfth feature map is processed by the fifth cross-stage module to output a thirteenth feature map; the thirteenth feature map is processed by the sixth convolutional layer regularization layer to output a fourteenth feature map; the fourteenth feature map is processed by the second upsampling layer to output a fifteenth feature map; the fifteenth feature map is processed by the second convolutional layer to output the first recognition result; wherein the target unlabeled image is any one of the multiple unlabeled images.

[0093] Specifically, the first convolution layer performs a convolution operation on the target unlabeled image through a first convolution kernel to capture local features in the target unlabeled image and generate the first feature map.

[0094] The first convolutional layer regularization layer is a structural unit, such as Figure 4As shown, the first convolutional layer regularization layer includes a fifth convolutional layer, a first regularization layer and a first activation function layer. The first regularization layer is a batch normalization layer. The first activation function layer is a Leaky ReLU (Leaky Rectified Linear Unit, with leaky linear rectifier function) activation function layer. In an embodiment of the present application, the fifth convolutional layer performs local feature extraction on the first feature map to obtain the second feature map. The first regularization layer is used to standardize and stabilize the training process and improve the performance of the first convolutional layer regularization layer. The first activation function layer constructs an efficient and stable convolutional neural network by increasing nonlinear capabilities and alleviating the gradient vanishing problem (that is, promoting gradient flow) to improve the expression ability and classification accuracy of the first convolutional layer regularization layer for image features. In addition, the first activation function layer is also used to solve the problem of neuron death. The fifth convolutional layer, the first regularization layer and the first activation function layer work together to improve the recognition accuracy of the second feature map.

[0095] The use of the regularization layer of the first convolutional layer is the application of the regularization method. The embodiment of the present application adopts the regularization method to learn a representation that is invariant across data domains and is robust to the differences between the source data domain and the target data domain.

[0096] The first cross-stage module is an architectural design for improving the performance and efficiency of convolutional neural networks. The first cross-stage module includes two hyperparameters, namely, a first hyperparameter and a second hyperparameter, wherein the first hyperparameter is represented by the presence of multiple first convolutional layer regularization layers in the first cross-stage module, Figure 5 As shown, in Figure 5 , the first cross-stage module is represented as a cross-stage module, the first convolutional layer regularization layer is represented as a convolutional layer regularization layer; the second hyperparameter is represented as the presence of multiple residual modules in the first cross-stage module. In the embodiment of the present application, the first hyperparameter and the second hyperparameter are both set to 1.

[0097] In some embodiments, Figure 5 As shown, the first cross-stage module also includes a second regularization layer and a second activation function layer. Figure 5 In , the second regularization layer is represented as a regularization layer, and the second activation function layer is represented as an activation function layer.

[0098] In some embodiments, the first cross-stage module divides the second feature map into two parts, one part of which directly skips the multiple first convolutional layer regularization layers of the first cross-stage module, and the other part is processed by the multiple first convolutional layer regularization layers (that is, multiple convolutional layers and activation functions) of the first cross-stage module and then fused with the previous part of the feature map, and then input into the second regularization layer and the second activation function layer and one first convolutional layer regularization layer in sequence for processing. This design helps to reduce the amount of calculation, balance the gradient flow of the first cross-stage module, and alleviate the problem of gradient disappearance or explosion, thereby improving the overall learning ability and generalization performance of the first target detection module, and then improving the accuracy of the first recognition result.

[0099] The structures of the second convolutional layer regularization layer, the third convolutional layer regularization layer, the fourth convolutional layer regularization layer, the fifth convolutional layer regularization layer and the sixth convolutional layer regularization layer are the same as the structures of the first convolutional layer regularization layer, and are not repeated here. The second convolutional layer regularization layer, the third convolutional layer regularization layer, the fourth convolutional layer regularization layer, the fifth convolutional layer regularization layer and the sixth convolutional layer regularization layer are all used to extract local features of the input image to obtain the corresponding feature map. The setting of the second convolutional layer regularization layer, the third convolutional layer regularization layer, the fourth convolutional layer regularization layer, the fifth convolutional layer regularization layer and the sixth convolutional layer regularization layer can further enhance the overall robustness of the first target detection module and further improve the accuracy of identifying the first recognition result.

[0100] The structures of the second cross-stage module, the third cross-stage module, the fourth cross-stage module and the fifth cross-stage module are the same as those of the first cross-stage module, so they are not repeated here. The second cross-stage module, the third cross-stage module, the fourth cross-stage module and the fifth cross-stage module are all used to further improve the overall learning ability and generalization performance of the first target detection module, thereby improving the accuracy of the first recognition result.

[0101] The first pooling layer is used to perform dimensionality reduction processing on the eighth feature map to reduce the amount of calculation, improve the overall processing speed of the first target detection module, and help prevent overfitting.

[0102] The first upsampling layer is used to improve the resolution of the eleventh feature map and process the eleventh feature map into a high-resolution twelfth feature map to improve image quality, thereby further improving the accuracy of recognition of the first recognition result.

[0103] The function of the second upsampling layer is the same as that of the first upsampling layer.

[0104] The function of the second convolutional layer is the same as that of the first convolutional layer.

[0105] In some implementations, the training method of the first target detection module includes:

[0106] S31, obtaining a first training data set and the first training pseudo-label set, wherein each first training data in the first training data set has a one-to-one correspondence with one of the first training pseudo-labels in the first training pseudo-label set.

[0107] Each first training data is a kind of unlabeled image data.

[0108] S32: Input the target first training data into the first initial detection module to obtain a first training result, wherein the target first training data is any first training data in the first training data set.

[0109] S33: Perform a non-maximum suppression operation on the first training result to obtain a third pseudo label.

[0110] S34. Calculate a first loss function based on a first training pseudo label corresponding to the target first training data and the third pseudo label.

[0111] S35. Adjust parameters of the first initial detection module based on the first loss function to form the first target detection module.

[0112] S36: All the third pseudo labels constitute a second training pseudo label set.

[0113] In some embodiments, the first initial detection module includes a first initial convolutional layer, a first initial convolutional layer regularization layer, a first initial cross-stage module, a second initial convolutional layer regularization layer, a second initial cross-stage module, a third initial convolutional layer regularization layer, a third initial cross-stage module, a fourth initial convolutional layer regularization layer, a first initial pooling layer, a fourth initial cross-stage module, a fifth initial convolutional layer regularization layer, a first initial upsampling layer, a fifth initial cross-stage module, a sixth initial convolutional layer regularization layer, a second initial upsampling layer, and a second initial convolutional layer, wherein the first initial convolutional layer The first convolution layer processes the target first training data and outputs a first initial feature map; the first initial convolution layer regularization layer processes the first initial feature map and outputs a second initial feature map; the first initial cross-stage module processes the second initial feature map and outputs a third initial feature map; the second initial convolution layer regularization layer processes the third initial feature map and outputs a fourth initial feature map; the second initial cross-stage module processes the fourth initial feature map and outputs a fifth initial feature map; the third initial convolution layer regularization layer processes the fifth initial feature map The first initial feature map is processed by the first initial convolution layer regularization layer to output a sixth initial feature map; the third initial cross-stage module processes the sixth initial feature map and outputs a seventh initial feature map; the fourth initial convolution layer regularization layer processes the seventh initial feature map and outputs an eighth initial feature map; the first initial pooling layer processes the eighth initial feature map and outputs a ninth initial feature map; the fourth initial cross-stage module processes the ninth initial feature map and outputs a tenth initial feature map; the fifth initial convolution layer regularization layer processes the tenth initial feature map and outputs an eleventh initial feature map; the first initial upsampling layer processes the eleventh initial feature map and outputs a twelfth initial feature map; the fifth initial cross-stage module processes the twelfth initial feature map and outputs a thirteenth initial feature map; the sixth initial convolution layer regularization layer processes the thirteenth initial feature map and outputs a fourteenth initial feature map; the second initial upsampling layer processes the fourteenth initial feature map and outputs a fifteenth initial feature map; the second initial convolution layer processes the fifteenth initial feature map and outputs the first training result.

[0114] It should be understood that the first initial convolutional layer has the same structure and function as the first convolutional layer, and the only difference between the two is the parameters, that is, the first initial convolutional layer can be obtained after being adjusted by the first loss function and / or the first cross pseudo-supervision loss function.

[0115] The first initial convolutional layer regularization layer has the same structure and function as the first convolutional layer regularization layer, and the only difference between the two is the parameters. That is, after the first initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross pseudo-supervised loss function, the first convolutional layer regularization layer can be obtained.

[0116] The first initial cross-stage module has the same structure and function as the first cross-stage module, and the only difference between the two is the parameters. That is, after the first initial cross-stage module is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the first cross-stage module can be obtained.

[0117] The second initial convolutional layer regularization layer and the second convolutional layer regularization layer have the same structure and function, and the only difference between the two is the parameters. That is, after the second initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross pseudo-supervised loss function, the second convolutional layer regularization layer can be obtained.

[0118] The second initial cross-stage module has the same structure and function as the second cross-stage module, and the only difference between the two is the parameters. That is, after the second initial cross-stage module is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the second cross-stage module can be obtained.

[0119] The third initial convolutional layer regularization layer has the same structure and function as the third convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the third initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the third convolutional layer regularization layer can be obtained.

[0120] The third initial cross-stage module has the same structure and function as the third cross-stage module, and the only difference between the two is the parameters. That is, after the third initial cross-stage module is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the third cross-stage module can be obtained.

[0121] The fourth initial convolutional layer regularization layer has the same structure and function as the fourth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the fourth initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the fourth convolutional layer regularization layer can be obtained.

[0122] The first initial pooling layer has the same structure and function as the first pooling layer, and the only difference between the two is the parameters. That is, the first initial pooling layer can be obtained after being adjusted by the first loss function and / or the first cross-pseudo-supervised loss function.

[0123] The fourth initial cross-stage module has the same structure and function as the fourth cross-stage module, and the only difference between the two is the parameters. That is, after the fourth initial cross-stage module is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the fourth cross-stage module can be obtained.

[0124] The fifth initial convolutional layer regularization layer has the same structure and function as the fifth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the fifth initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the fifth convolutional layer regularization layer can be obtained.

[0125] The first initial upsampling layer has the same structure and function as the first upsampling layer, and the only difference between the two is the parameters. That is, the first initial upsampling layer can be obtained after being adjusted by the first loss function and / or the first cross pseudo-supervised loss function.

[0126] The fifth initial cross-stage module has the same structure and function as the fifth cross-stage module, and the only difference between the two is the parameters. That is, after the fifth initial cross-stage module is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the fifth cross-stage module can be obtained.

[0127] The sixth initial convolutional layer regularization layer has the same structure and function as the sixth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the sixth initial convolutional layer regularization layer is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the sixth convolutional layer regularization layer can be obtained.

[0128] The second initial upsampling layer and the second upsampling layer have the same structure and function, and the only difference between the two is the parameters. That is, after the second initial upsampling layer is adjusted by the first loss function and / or the first cross-pseudo-supervised loss function, the second upsampling layer can be obtained.

[0129] The second initial convolutional layer and the second convolutional layer have the same structure and function, and the only difference between the two is the parameters. That is, the second initial convolutional layer can be obtained after being adjusted by the first loss function and / or the first cross pseudo-supervision loss function.

[0130] In some implementations, after step S2, the small target recognition method further includes:

[0131] A non-maximum suppression operation is performed on the first recognition result to obtain a fifth pseudo label, and the fifth pseudo label is added to a second training pseudo label set, wherein the second training pseudo label set is used to train a second target detection module.

[0132] In some implementations, the small target recognition method further includes:

[0133] S4. Acquire multiple labeled images.

[0134] S5. Use a second target detection module to perform small target recognition on each of the marked images and output a second recognition result.

[0135] S6. Perform a non-maximum suppression operation on the second recognition result to obtain a first pseudo label, and add the first pseudo label to a first training pseudo label set, wherein the first training pseudo label set is used to train the first target detection module.

[0136] like Figure 3 As shown, the second target detection module includes a third convolutional layer, a seventh convolutional layer regularization layer, a sixth cross-stage module, an eighth convolutional layer regularization layer, a seventh cross-stage module, a ninth convolutional layer regularization layer, an eighth cross-stage module, a tenth convolutional layer regularization layer, a second pooling layer, a ninth cross-stage module, an eleventh convolutional layer regularization layer, a third upsampling layer, a tenth cross-stage module, a twelfth convolutional layer regularization layer, a fourth upsampling layer and a fourth convolutional layer.

[0137] Further, the step of using the second target detection module to perform small target recognition on each of the labeled images and outputting the second recognition result includes: using the third convolutional layer to process the target labeled image and outputting the sixteenth feature map; using the seventh convolutional layer regularization layer to process the sixteenth feature map and outputting the seventeenth feature map; using the sixth cross-stage module to process the seventeenth feature map and outputting the eighteenth feature map; using the eighth convolutional layer regularization layer to process the eighteenth feature map and outputting the nineteenth feature map; using the seventh cross-stage module to process the nineteenth feature map and outputting the twentieth feature map; using the ninth convolutional layer regularization layer to process the twentieth feature map and outputting the twenty-first feature map; using the eighth cross-stage module to process the twenty-first feature map and outputting the twenty-second feature map; using the tenth convolutional layer regularization layer to process the twenty-second feature map and output the second thirteenth feature map; the second pooling layer is used to process the twenty-third feature map, and the twenty-fourth feature map is output; the ninth cross-stage module is used to process the twenty-fourth feature map, and the twenty-fifth feature map is output; the eleventh convolutional layer regularization layer is used to process the twenty-fifth feature map, and the twenty-sixth feature map is output; the third upsampling layer is used to process the twenty-sixth feature map, and the twenty-seventh feature map is output; the tenth cross-stage module is used to process the twenty-seventh feature map, and the twenty-eighth feature map is output; the twelfth convolutional layer regularization layer is used to process the twenty-eighth feature map, and the twenty-ninth feature map is output; the fourth upsampling layer is used to process the twenty-ninth feature map, and the thirtieth feature map is output; the fourth convolutional layer is used to process the thirtieth feature map, and the second recognition result is output; wherein the target labeled image is any one of the multiple labeled images. It should be understood that, Figure 3 The dotted arrows in denoted as loss supervision.

[0138] The third convolutional layer has the same structure and function as the first convolutional layer, and the only difference between the two is the initialization.

[0139] The seventh convolutional layer regularization layer has the same structure and function as the first convolutional layer regularization layer, and the only difference between the two is the initialization.

[0140] The sixth cross-stage module has the same structure and function as the first cross-stage module, and the only difference between the two is the initialization.

[0141] The eighth convolutional layer regularization layer has the same structure and function as the second convolutional layer regularization layer, and the only difference between the two is the initialization.

[0142] The seventh cross-stage module has the same structure and function as the second cross-stage module, and the only difference between the two is the initialization.

[0143] The ninth convolutional layer regularization layer has the same structure and function as the third convolutional layer regularization layer, and the only difference between the two is the initialization.

[0144] The eighth cross-stage module has the same structure and function as the third cross-stage module, and the only difference between the two is the initialization.

[0145] The tenth convolutional layer regularization layer has the same structure and function as the fourth convolutional layer regularization layer, and the only difference between the two is the initialization.

[0146] The second pooling layer has the same structure and function as the first pooling layer, and the only difference between the two is the initialization.

[0147] The ninth cross-stage module has the same structure and function as the fourth cross-stage module, and the only difference between the two is the initialization.

[0148] The eleventh convolutional layer regularization layer has the same structure and function as the fifth convolutional layer regularization layer, and the only difference between the two is the initialization.

[0149] The third upsampling layer has the same structure and function as the first upsampling layer, and the only difference between the two is the initialization.

[0150] The tenth cross-stage module has the same structure and function as the fifth cross-stage module, and the only difference between the two is the initialization.

[0151] The twelfth convolutional layer regularization layer and the sixth convolutional layer regularization layer have the same structure and function, and the only difference between the two is the initialization.

[0152] The fourth upsampling layer has the same structure and function as the second upsampling layer, and the only difference between the two is the initialization.

[0153] The fourth convolutional layer has the same structure and function as the second convolutional layer, and the only difference between the two is the initialization.

[0154] In some implementations, the training method of the second object detection module includes:

[0155] S61. Acquire a second training data set and the second training pseudo label set, wherein each second training data in the second training data set has a one-to-one correspondence with one of the second training pseudo labels in the second training pseudo label set.

[0156] Each second training data is a type of labeled image data.

[0157] S62: Input the target second training data into the second initial detection module to obtain a second training result, wherein the target second training data is any second training data in the second training data set.

[0158] S63: Perform a non-maximum suppression operation on the second training result to obtain a fourth pseudo label.

[0159] S64: Calculate a second loss function based on a second training pseudo label corresponding to the target second training data and the fourth pseudo label.

[0160] S65. Adjust parameters of the second initial detection module based on the second loss function to form the second target detection module.

[0161] S66: Add all the fourth pseudo labels to the first training pseudo label set.

[0162] Further, the second initial detection module includes a third initial convolution layer, a seventh initial convolution layer regularization layer, a sixth initial cross-stage module, an eighth initial convolution layer regularization layer, a seventh initial cross-stage module, a ninth initial convolution layer regularization layer, an eighth initial cross-stage module, a tenth initial convolution layer regularization layer, a second initial pooling layer, a ninth initial cross-stage module, an eleventh initial convolution layer regularization layer, a third initial upsampling layer, a tenth initial cross-stage module, a twelfth initial convolution layer regularization layer, a fourth initial upsampling layer and a fourth initial convolution layer; wherein the third initial convolution layer performs a first-order pooling operation on the target second training data. The seventh initial convolutional layer regularization layer processes the sixteenth initial feature map and outputs a seventeenth initial feature map; the sixth initial cross-stage module processes the seventeenth initial feature map and outputs an eighteenth initial feature map; the eighth initial convolutional layer regularization layer processes the eighteenth initial feature map and outputs a nineteenth initial feature map; the seventh initial cross-stage module processes the nineteenth initial feature map and outputs a twentieth initial feature map; the ninth initial convolutional layer regularization layer processes the twentieth initial feature map and outputs a Twenty-first initial feature map; the eighth initial cross-stage module processes the twenty-first initial feature map and outputs a twenty-second initial feature map; the tenth initial convolutional layer regularization layer processes the twenty-second initial feature map and outputs a twenty-third initial feature map; the second initial pooling layer processes the twenty-third initial feature map and outputs a twenty-fourth initial feature map; the ninth initial cross-stage module processes the twenty-fourth initial feature map and outputs a twenty-fifth initial feature map; the eleventh initial convolutional layer regularization layer processes the twenty-fifth initial feature map and outputs a twenty-sixth initial feature map; the third initial upsampling layer processes the twenty-sixth initial feature map and outputs a twenty-seventh initial feature map; the tenth initial cross-stage module processes the twenty-seventh initial feature map and outputs a twenty-eighth initial feature map; the twelfth initial convolutional layer regularization layer processes the twenty-eighth initial feature map and outputs a twenty-ninth initial feature map; the fourth initial upsampling layer processes the twenty-ninth initial feature map and outputs the second training result.

[0163] It should be understood that the third initial convolutional layer has the same structure and function as the third convolutional layer, and the only difference between the two is the parameters, that is, the third initial convolutional layer can be obtained after being adjusted by the second loss function and / or the second cross-pseudo-supervised loss function.

[0164] The seventh initial convolutional layer regularization layer has the same structure and function as the seventh convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the seventh initial convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the seventh convolutional layer regularization layer can be obtained.

[0165] The sixth initial cross-stage module has the same structure and function as the sixth cross-stage module, and the only difference between the two is the parameters. That is, after the sixth initial cross-stage module is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the sixth cross-stage module can be obtained.

[0166] The eighth initial convolutional layer regularization layer has the same structure and function as the eighth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the eighth initial convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the eighth convolutional layer regularization layer can be obtained.

[0167] The seventh initial cross-stage module has the same structure and function as the seventh cross-stage module, and the only difference between the two is the parameters. That is, after the seventh initial cross-stage module is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the seventh cross-stage module can be obtained.

[0168] The ninth initial convolutional layer regularization layer and the ninth convolutional layer regularization layer have the same structure and function, and the only difference between the two is the parameters, that is, after the ninth convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the ninth convolutional layer regularization layer can be obtained.

[0169] The eighth initial cross-stage module has the same structure and function as the eighth cross-stage module, and the only difference between the two is the parameters. That is, after the eighth initial cross-stage module is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the eighth cross-stage module can be obtained.

[0170] The tenth initial convolutional layer regularization layer has the same structure and function as the tenth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the tenth initial convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the tenth convolutional layer regularization layer can be obtained.

[0171] The second initial pooling layer has the same structure and function as the second pooling layer, and the only difference between the two is the parameters. That is, the second initial pooling layer can be obtained after the second loss function and / or the second cross-pseudo-supervised loss function is adjusted.

[0172] The ninth initial cross-stage module has the same structure and function as the ninth cross-stage module, and the only difference between the two is the parameters. That is, after the ninth initial cross-stage module is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the ninth cross-stage module can be obtained.

[0173] The eleventh initial convolutional layer regularization layer has the same structure and function as the eleventh convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the eleventh initial convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the eleventh convolutional layer regularization layer can be obtained.

[0174] The third initial upsampling layer has the same structure and function as the third upsampling layer, and the only difference between the two is the parameters. That is, after the third initial upsampling layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the third upsampling layer can be obtained.

[0175] The tenth initial cross-stage module has the same structure and function as the tenth cross-stage module, and the only difference between the two is the parameters. That is, after the tenth initial cross-stage module is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the tenth cross-stage module can be obtained.

[0176] The twelfth initial convolutional layer regularization layer has the same structure and function as the twelfth convolutional layer regularization layer, and the only difference between the two is the parameters, that is, after the twelfth initial convolutional layer regularization layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the twelfth convolutional layer regularization layer can be obtained.

[0177] The fourth initial upsampling layer has the same structure and function as the fourth upsampling layer, and the only difference between the two is the parameters. That is, after the fourth initial upsampling layer is adjusted by the second loss function and / or the second cross-pseudo-supervised loss function, the fourth upsampling layer can be obtained.

[0178] The fourth initial convolutional layer has the same structure and function as the fourth convolutional layer, and the only difference between the two is the parameters, that is, the fourth initial convolutional layer can be obtained after being adjusted by the second loss function and / or the second cross pseudo-supervision loss function.

[0179] In some implementations, in order to promote cross-learning between the first target detection module and the second target detection module, the present application adds pseudo labels to the loss calculation method, and the specific symbols are represented as follows:

[0180] , ;

[0181] , ;

[0182] in, The first initial target detection module; The second initial target detection module; The input of the first initial target detection module is an unlabeled image that has undergone data regularization and enhancement operation; The input of the second initial target detection module is a labeled image that has undergone a data regularization enhancement operation; both the unlabeled image and the labeled image have undergone the same data regularization enhancement operation; is the output result of the first initial target detection module, that is, the first training result; is the output result of the second initial target detection module, that is, the second training result; It is a non-maximum suppression technology (Non Maximum Suppression); The pseudo labels (i.e., the third pseudo labels) generated by the first initial target detection module specifically include the pseudo labels of the types of feature points generated by the first initial target detection module, the pseudo labels of the prediction boxes generated by the first initial target detection module at the feature points, and the pseudo labels of whether there are objects at the feature points determined by the first initial target detection module; The pseudo label generated by the second initial target detection module (i.e., the fourth pseudo label) specifically includes the pseudo label of the feature point type by the second initial target detection module, the pseudo label of the prediction box at the feature point by the second initial target detection module, and the pseudo label of whether there is an object at the feature point by the second initial target detection module. The third pseudo label is a pseudo element of the second training pseudo label set, and the fourth pseudo label is a pseudo element of the first training pseudo label set.

[0183] In some embodiments, the calculation formula of the first loss function is:

[0184] ;

[0185] in, represents the cross entropy loss function; Represents the pseudo-label of the first initial target detection module for the type of feature point, that is, the pseudo-label of the type of feature point in the third pseudo-label; Represents the pseudo-label of the feature point type by the second initial target detection module, that is, the pseudo-label of the feature point type in the fourth pseudo-label; represents the CIOU calculation method; Represents the pseudo label of the predicted box at the feature point by the first initial target detection module, that is, the pseudo label of the predicted box at the feature point in the third pseudo label; Represents the pseudo label of the predicted box at the feature point by the second initial target detection module, that is, the pseudo label of the predicted box at the feature point in the fourth pseudo label; The pseudo label representing whether the first initial target detection module determines whether there is an object at the feature point, that is, the pseudo label representing whether there is an object at the feature point in the third pseudo label; It represents the pseudo label of whether the second initial target detection module determines whether there is an object at the feature point, that is, the pseudo label of whether the feature point is determined by the fourth pseudo label.

[0186] In some embodiments, the first loss function is a cross pseudo-supervisory loss function. The cross pseudo-supervisory loss function is bidirectional, one direction is , the other direction is In this application, the .

[0187] In some embodiments, the calculation formula of the second loss function is as follows:

[0188] ;

[0189] in, represents the cross entropy loss function; Represents the pseudo-label of the first initial target detection module for the type of feature point, that is, the pseudo-label of the type of feature point in the third pseudo-label; Represents the pseudo-label of the feature point type by the second initial target detection module, that is, the pseudo-label of the feature point type in the fourth pseudo-label; represents the CIOU calculation method, Represents the pseudo label of the predicted box at the feature point by the first initial target detection module, that is, the pseudo label of the predicted box at the feature point in the third pseudo label; Represents the pseudo label of the predicted box at the feature point by the second initial target detection module, that is, the pseudo label of the predicted box at the feature point in the fourth pseudo label; The pseudo label representing whether the first initial target detection module determines whether there is an object at the feature point, that is, the pseudo label representing whether there is an object at the feature point in the third pseudo label; It represents the pseudo label of whether the second initial target detection module determines whether there is an object at the feature point, that is, the pseudo label of whether the feature point is determined by the fourth pseudo label.

[0190] In some implementations, the training method of the first target detection module further includes:

[0191] S37. Generate a first cross-pseudo-supervised loss function based on the first loss function and the second loss function.

[0192] S38. Adjust the parameters of the first target detection module based on the first cross pseudo-supervision loss function.

[0193] Furthermore, the calculation formula of the first cross pseudo-supervision loss function is:

[0194] ;

[0195] in, represents the second loss function, represents the calculation coefficients of the first loss function and the second loss function, Represents the first loss function.

[0196] In some implementations, the training method of the second object detection module further includes:

[0197] S67. Generate a second cross pseudo-supervision loss function based on the first loss function and the second loss function.

[0198] S68. Adjust the parameters of the second target detection module based on the second cross pseudo-supervision loss function.

[0199] Furthermore, the calculation formula of the second cross pseudo-supervision loss function is the same as the calculation formula of the first cross pseudo-supervision loss function, so it is not repeated here.

[0200] In the above scheme, the small target recognition method provided in the embodiment of the present application designs a first target detection module, and the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image, so that each unlabeled image can obtain a clear first recognition result after being recognized by the first target detection module. Therefore, compared with the existing technology with poor recognition effect, the present application can improve the accuracy of small target recognition without determining the data domain of the unlabeled image.

[0201] The small target recognition method provided by the present application overcomes the defect that the data domain of unlabeled images is uncertain.

[0202] refer to Figure 6 As shown, a small target recognition system provided by the second aspect of an embodiment of the present application is provided. The small target recognition system 700 includes:

[0203] An acquisition module 701 is used to acquire a plurality of unmarked images;

[0204] The first recognition result output module 702 is used to use the first target detection module to perform small target recognition on each of the unlabeled images and output a first recognition result, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image.

[0205] The third aspect of the embodiments of the present application provides a terminal device, which includes a processor, a memory, a network interface, a display screen and a temperature sensor connected via a system bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a small target recognition method is implemented. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the temperature sensor is pre-set inside the terminal device to detect the operating temperature of the internal device.

[0206] In some embodiments, an embodiment of the present application provides a terminal device, which includes a processor and a memory, the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to execute the steps of the small target recognition method provided in the first aspect of the above-mentioned embodiment of the present application.

[0207] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, which is used to store a computer program, and the computer program enables a computer to execute the steps of the small target recognition method provided in the first aspect of the embodiment of the present application.

[0208] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0209] The technical features of the above embodiments can be combined without changing the basic principles of the present invention. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0210] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the scope of patent protection of the present application shall be subject to the attached claims.

Claims

1. A small target recognition method, characterized in that: include: Acquire multiple unlabeled images; Using a first target detection module to perform small target recognition on each of the unlabeled images, and outputting a first recognition result, wherein the first target detection module incorporates a regularization method to determine the data domain of the unlabeled image; The step of using the first target detection module to identify small targets in each of the unlabeled images and outputting a first identification result includes: using the first convolutional layer of the first target detection module to process the target unlabeled image and output a first feature map; using the first convolutional layer regularization layer of the first target detection module to process the first feature map and output a second feature map; using the first cross-stage module of the first target detection module to process the second feature map and output a third feature map; using the second convolutional layer regularization layer of the first target detection module to process the third feature map and output a fourth feature map; using the second cross-stage module of the first target detection module to process the fourth feature map and output a fifth feature map; using the third convolutional layer regularization layer of the first target detection module to process the fifth feature map and output a sixth feature map; using the third cross-stage module of the first target detection module to process the sixth feature map and output a seventh feature map; using the fourth convolutional layer regularization layer of the first target detection module to process the seventh feature map and output an eighth feature map. The first target detection module processes the eighth feature map by the first pooling layer and outputs the ninth feature map; the fourth cross-stage module of the first target detection module processes the ninth feature map and outputs the tenth feature map; the fifth convolutional layer regularization layer of the first target detection module processes the tenth feature map and outputs the eleventh feature map; the first upsampling layer of the first target detection module processes the eleventh feature map and outputs the twelfth feature map; the fifth cross-stage module of the first target detection module processes the twelfth feature map and outputs the thirteenth feature map; the sixth convolutional layer regularization layer of the first target detection module processes the thirteenth feature map and outputs the fourteenth feature map; the second upsampling layer of the first target detection module processes the fourteenth feature map and outputs the fifteenth feature map; the second convolutional layer of the first target detection module processes the fifteenth feature map and outputs the first recognition result; wherein the target unlabeled image is any one of the multiple unlabeled images; The training method of the first target detection module includes: obtaining a first training data set and a first training pseudo-label set, wherein each first training data in the first training data set has a one-to-one correspondence with one of the first training pseudo-labels in the first training pseudo-label set; inputting the target first training data into the first initial detection module to obtain a first training result, wherein the target first training data is any first training data in the first training data set; performing a non-maximum suppression operation on the first training result to obtain a third pseudo-label; calculating a first loss function based on the first training pseudo-label corresponding to the target first training data and the third pseudo-label; adjusting the parameters of the first initial detection module based on the first loss function to form the first target detection module; and all the third pseudo-labels constitute a second training pseudo-label set.

2. The small target recognition method according to claim 1, characterized in that: The small target recognition method also includes: Obtain multiple labeled images; Using a second target detection module to perform small target recognition on each of the marked images, and output a second recognition result; A non-maximum suppression operation is performed on the second recognition result to obtain a first pseudo label, and the first pseudo label is added to the first training pseudo label set, wherein the first training pseudo label set is used to train the first target detection module.

3. The small target recognition method according to claim 2, characterized in that: The step of using the second target detection module to perform small target recognition on each of the marked images and outputting a second recognition result comprises: Processing the target labeled image using the third convolutional layer of the second target detection module to output a sixteenth feature map; Processing the sixteenth feature map using the seventh convolutional layer regularization layer of the second object detection module, and outputting a seventeenth feature map; Using the sixth cross-stage module of the second object detection module to process the seventeenth feature map, and output an eighteenth feature map; Using the eighth convolutional layer regularization layer of the second object detection module to process the eighteenth feature map, and output a nineteenth feature map; Using the seventh cross-stage module of the second object detection module to process the nineteenth feature map, and output a twentieth feature map; Using the ninth convolutional layer regularization layer of the second object detection module to process the 20th feature map, and output a 21st feature map; Using the eighth cross-stage module of the second object detection module to process the twenty-first feature map, and output a twenty-second feature map; Using the tenth convolutional layer regularization layer of the second object detection module to process the twenty-second feature map, and output a twenty-third feature map; Processing the twenty-third feature map using the second pooling layer of the second object detection module, and outputting a twenty-fourth feature map; Using a ninth cross-stage module of the second object detection module to process the twenty-fourth feature map, and output a twenty-fifth feature map; Using the eleventh convolutional layer regularization layer of the second object detection module to process the twenty-fifth feature map, and output a twenty-sixth feature map; Processing the twenty-sixth feature map using the third upsampling layer of the second object detection module, and outputting a twenty-seventh feature map; Processing the twenty-seventh feature map using a tenth cross-stage module of the second object detection module to output a twenty-eighth feature map; Processing the twenty-eighth feature map using the twelfth convolutional layer regularization layer of the second object detection module, and outputting a twenty-ninth feature map; Processing the twenty-ninth feature map using the fourth upsampling layer of the second object detection module, and outputting a thirtieth feature map; Processing the thirtieth feature map using the fourth convolutional layer of the second object detection module, and outputting the second recognition result; The target labeled image is any labeled image among the multiple labeled images.

4. The small target recognition method according to claim 2, characterized in that: The training method of the second target detection module includes: Acquire a second training data set and the second training pseudo-label set, wherein each second training data in the second training data set has a one-to-one correspondence with one of the second training pseudo-labels in the second training pseudo-label set; Inputting the target second training data into the second initial detection module to obtain a second training result, wherein the target second training data is any second training data in the second training data set; Performing a non-maximum suppression operation on the second training result to obtain a fourth pseudo label; Calculate a second loss function based on a second training pseudo label corresponding to the target second training data and the fourth pseudo label; adjusting parameters of the second initial detection module based on the second loss function to form the second target detection module; All the fourth pseudo labels are added to the first training pseudo label set.

5. The small target recognition method according to claim 4, characterized in that: The training method of the first target detection module also includes: generating a first cross pseudo-supervisory loss function based on the first loss function and the second loss function; adjusting parameters of the first target detection module based on the first cross pseudo-supervisory loss function; The training method of the second target detection module also includes: generating a second cross-pseudo-supervised loss function based on the first loss function and the second loss function; and adjusting the parameters of the second target detection module based on the second cross-pseudo-supervised loss function.

6. A small target recognition system, characterized in that: include: An acquisition module, used for acquiring a plurality of unlabeled images; A first recognition result output module, configured to use a first target detection module to perform small target recognition on each of the unlabeled images and output a first recognition result, wherein the first target detection module incorporates a regularization method to determine a data domain of the unlabeled image; The small target recognition system is also used to use the first convolutional layer of the first target detection module to process the target unlabeled image and output a first feature map; use the first convolutional layer regularization layer of the first target detection module to process the first feature map and output a second feature map; use the first cross-stage module of the first target detection module to process the second feature map and output a third feature map; use the second convolutional layer regularization layer of the first target detection module to process the third feature map and output a fourth feature map; use the second cross-stage module of the first target detection module to process the fourth feature map and output a fifth feature map; use the third convolutional layer regularization layer of the first target detection module to process the fifth feature map and output a sixth feature map; use the third cross-stage module of the first target detection module to process the sixth feature map and output a seventh feature map; use the fourth convolutional layer regularization layer of the first target detection module to process the seventh feature map and output an eighth feature map; use the first target detection module The first pooling layer processes the eighth feature map and outputs a ninth feature map; the fourth cross-stage module of the first target detection module is used to process the ninth feature map and output a tenth feature map; the fifth convolutional layer regularization layer of the first target detection module is used to process the tenth feature map and output an eleventh feature map; the first upsampling layer of the first target detection module is used to process the eleventh feature map and output a twelfth feature map; the fifth cross-stage module of the first target detection module is used to process the twelfth feature map and output a thirteenth feature map; the sixth convolutional layer regularization layer of the first target detection module is used to process the thirteenth feature map and output a fourteenth feature map; the second upsampling layer of the first target detection module is used to process the fourteenth feature map and output a fifteenth feature map; the second convolutional layer of the first target detection module is used to process the fifteenth feature map and output the first recognition result; wherein the target unlabeled image is any one of the multiple unlabeled images; The small target recognition system is also used to obtain a first training data set and a first training pseudo-label set, wherein each first training data in the first training data set has a one-to-one correspondence with one of the first training pseudo-labels in the first training pseudo-label set; input the target first training data into a first initial detection module to obtain a first training result, wherein the target first training data is any first training data in the first training data set; perform a non-maximum suppression operation on the first training result to obtain a third pseudo-label; calculate a first loss function based on the first training pseudo-label corresponding to the target first training data and the third pseudo-label; adjust the parameters of the first initial detection module based on the first loss function to form the first target detection module; and constitute all the third pseudo-labels into a second training pseudo-label set.

7. A terminal device, characterized in that: include: A processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory to execute the steps of the small target recognition method described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein the computer program enables a computer to execute the steps of the small target recognition method according to any one of claims 1 to 5.

Images