A target detection method, device and storage medium thereof

Abstract

The present invention discloses a target detection method, device and storage medium thereof, the method comprising receiving a labeled model image and an unlabeled model image, training the labeled model image and the unlabeled model image according to an average teacher strategy through a pre-built student network and a teacher network to obtain a labeled model image loss and an unlabeled model image loss; obtaining a training model according to the labeled model image loss and the unlabeled model image loss; receiving an image to be identified; and detecting the image to be identified using the training model. The present application trains the labeled model image and the unlabeled model image according to the average teacher strategy through a neural network to obtain a training model, and uses the training model to detect images, which can greatly improve the accuracy of recognition detection.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to a target detection method, a device and a storage medium thereof. Background Art

[0002] In recent years, the visual research on generating predictive labels by synthesizing data based on deep learning models has attracted more and more attention. For example, in some images, it is necessary to identify the target objects such as cars and people in the image, and use labels to locate and annotate the appropriate range outside the target object. Although deep neural networks have been proven to be an efficient method for learning visual models under large data sets, there are still widespread errors when simply using the learning model obtained by deep neural networks to synthesize images to generate predictive labels.

[0003] Therefore, the inventors have discovered that there are at least the following problems in the prior art: when simply using a learning model obtained by a deep neural network to generate predictive labels for identifying objects, there are still common technical problems of errors. Summary of the invention

[0004] The present application provides a target detection method, which adopts an average teacher strategy to train labeled model images and unlabeled models, and finally generates a training model, so as to overcome the common errors that still exist when simply using a learning model obtained by a deep neural network to generate predictive labels to identify objects.

[0005] The method specifically includes:

[0006] Receiving a labeled model image and an unlabeled model image, wherein the labeled model image is an image with a pre-built recognition object classification data set, and the unlabeled model image is an image without a pre-built recognition object classification data set;

[0007] The labeled model image and the unlabeled model image are trained by a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss;

[0008] A training model is obtained by calculating the loss of the labeled model image and the loss of the unlabeled model image;

[0009] Receiving an image to be recognized;

[0010] The training model is used to detect the image to be identified to obtain a detection result of the identified object.

[0011] Optionally, training the labeled model image and the unlabeled model image through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss value and an unlabeled model image loss value includes:

[0012] Inputting the labeled model image into the student network for training to obtain the labeled model image loss;

[0013] Performing two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image;

[0014] Inputting the first noise interference image into the teacher network training to obtain a first similarity feature and a candidate region of the teacher network;

[0015] Input the candidate region and the second noise interference image into the student network for training to obtain a second similarity feature;

[0016] The unlabeled model image loss is calculated according to the first similarity feature and the second similarity feature.

[0017] Optionally, the student network and the teacher network are Faster-R-CNN networks;

[0018] The Faster-R-CNN network has a Base CNN network, an RCNN network, and an RPN network;

[0019] Optionally, after the step of calculating the training model according to the loss value of the labeled model image and the loss value of the unlabeled model image, the method further comprises:

[0020] Adjusting the weight of the unlabeled model image loss;

[0021] A training model is obtained by calculating the loss of the labeled model image and the adjusted loss of the unlabeled model image.

[0022] Optionally, inputting the labeled model image into a student network for training to obtain a labeled model image loss includes a training step of optimizing the student network by a stochastic gradient descent algorithm.

[0023] The present application also provides a target detection device, which includes:

[0024] A receiving module, used for receiving annotated model images and unlabeled model images, wherein the annotated model images are image sets with pre-built recognition object classification data sets, and the unlabeled model images are image sets without pre-built recognition object classification data sets;

[0025] A training module, used for training the labeled model image and the unlabeled model image through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss;

[0026] A calculation module, used for calculating a training model according to the labeled model image loss and the unlabeled model image loss;

[0027] The receiving module is further used to receive the image to be identified;

[0028] The recognition module is used to detect the image to be recognized using the training model to obtain a detection result of the recognized object.

[0029] Optionally, the device comprises:

[0030] The training module is further used to input the labeled model image into the student network for training to obtain the labeled model image loss;

[0031] A noise module, used for performing two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image;

[0032] The training module is further used to perform two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image; and is further used to input the candidate area and the second noise interference image into the student network for training to obtain a second similarity feature;

[0033] The calculation module is further used to obtain an unlabeled model image loss through training based on the first similarity feature and the second similarity feature.

[0034] Optionally, the device comprises:

[0035] An adjustment module, used for adjusting the weight of the unlabeled model image loss;

[0036] The calculation module is also used to calculate the training model according to the labeled model image loss and the adjusted unlabeled model image loss.

[0037] The present application provides a target detection device, the device comprising: a processor and a memory;

[0038] The memory stores an application program executable by the processor, which is used to enable the processor to execute the steps of the target detection method.

[0039] A computer-readable storage medium stores a computer program, wherein the program implements the steps of the target detection method when executed by a processor.

[0040] As can be seen above, based on the above embodiments, the present application complies with the semi-supervised learning method and can achieve adaptive annotation training results. The training model calculated by the labeled model image loss and the unlabeled model image loss can greatly improve the prediction accuracy of the detection of the received image to be identified through the training model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of a target detection method process 100 in one embodiment of the present application;

[0042] Figure 2 A schematic diagram of a target detection method process 200 in one embodiment of the present application;

[0043] Figure 3 This is a schematic diagram of a target detection method process 300 in one embodiment of the present application;

[0044] Figure 4 This is a schematic diagram of a target detection method process 400 in one embodiment of the present application;

[0045] Figure 5 A schematic diagram of a target detection method process 500 in one embodiment of the present application;

[0046] Figure 6 This is a schematic diagram of comparing the target detection accuracy after the weight coefficient is changed in one embodiment of the present application;

[0047] Figure 7 This is a schematic diagram of comparing the target detection accuracy after the smoothing coefficient is changed in one embodiment of the present application;

[0048] Figure 8 It is a schematic diagram for comparing the target detection accuracy of different methods after the weight coefficient and smoothing coefficient are changed;

[0049] Fig. 9 A schematic diagram of an experimental comparison of target detection using different methods;

[0050] Fig.10 This is another schematic diagram of experimental comparison of target detection accuracy using different methods;

[0051] Fig.11 This is a schematic diagram of the architecture of a target detection device in one embodiment of the present application. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0053] In some embodiments of the present application, in order to solve the problems raised in the background technology, for target detection after scene change, and predictive labeling of identified target objects, an average teacher solution is introduced, that is, a teacher network and a student network are established for image processing, and pre-adaptation is changed to semi-supervised learning. Unsupervised learning refers to the fact that the labeling information of training samples is unknown. The goal is to reveal the intrinsic properties and laws of the data by learning from unlabeled training samples, providing a basis for further data analysis. Semi-supervised learning refers to the fact that the training set contains both labeled sample data and unlabeled sample data, without the need for human intervention, so that machine learning does not rely on external interaction and automatically uses unlabeled samples to improve learning performance. From the above descriptions of unsupervised learning and semi-supervised learning, it can be seen that unsupervised learning requires human intervention, which increases costs, while the semi-supervised learning of the present application does not require human intervention.

[0054] The problem caused by the scene change can be naturally bridged by the uniform loss of the teacher network and the student network. In this application, the improved average teacher scheme is applied to the target detection of scene change. This application is particularly creative to provide a target detection method based on the Faster R-CNN framework that integrates the uniform measurement of the student network and the teacher network.

[0055] Theoretically, the above-mentioned target detection method first learns the similarities captured in two related images in different scenes through the teacher network and the student network respectively. The overall architecture of this application is to optimize the three consistency rules.

[0056] Among them, the first item is to unify the prediction results between the image scenes to be identified provided by the teacher network and the student network scenes, referred to as inter-regional consistency;

[0057] The second item is to unify the structure of the image to be identified that matches the teacher network and the student network, referred to as the consistency between the images;

[0058] The third item is to unify the enhanced similarity of similar regions in similar scenes in the student network, referred to as unification in the figure.

[0059] Figure 1 FIG. 1 is a schematic diagram of a target detection method process 100 in one embodiment of the present application. Figure 1 As shown, in one embodiment, the present application provides a target detection method, the method comprising:

[0060] S101, receiving a labeled model image and an unlabeled model image, wherein the labeled model image is an image with a pre-built recognition object classification data set, and the unlabeled model image is an image without a pre-built recognition object classification data set;

[0061] The labeled model image in this step can be understood as the image of the identified object annotated by the label, such as Figure 4 As shown in the figure, the image on the lower left can be understood as a labeled model image, that is, plants, cars, people, etc. have been labeled with different labels, and the corresponding image on the lower right can be understood as an unlabeled model image, because there are no labels in the image. It should be pointed out that the concept of classified datasets can be understood as labeling by labeling. The dataset can be set as follows: the red label (the box in the figure) is classified as a person, so as to define and create a dataset.

[0062] S102, training the labeled model image and the unlabeled model image through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss;

[0063] What needs to be pointed out in this step is the average teacher strategy (i.e. Mean Teacher). Specifically, the average teacher strategy provides a student network and a teacher network (the neural network structure of the two is the same, so the student network or the teacher network is defined according to the parameter update method), the student network is used to process the labeled model image, and the teacher network is used to process the unlabeled model image, and then the teacher network can guide the student network to achieve the effect of having common attributes. Therefore, it should be particularly pointed out that there are common attributes between the unlabeled model image losses obtained by training according to the average teacher strategy in this application, and the common attributes here can be used to improve the prediction accuracy of recognition, and the specific content will be described in detail later, so it will not be repeated here.

[0064] Figure 2 FIG. 2 is a schematic diagram of a target detection method process 200 in one embodiment of the present application. Figure 2 As shown, the target detector directly applies a neural network to a real image to be identified (lower left corner) by training the training model obtained by training the 3D CAD model synthetic image (upper left corner) to perform target detection and labeling (upper right corner, i.e., the prior art) and uses the method of the present application (i.e., the average teacher strategy) to perform target detection and labeling (lower right corner) The process results are compared.

[0065] In the prior art, a training model based on 3D CAD synthetic data cannot accurately locate objects such as people and cars in the recognition image. In the present application, accurate results are obtained under the method provided in the present application.

[0066] S103, obtaining a training model by calculating the loss of the labeled model image and the loss of the unlabeled model image;

[0067] We can use the formula to explain this step:

[0068] Define the labeled model image as:

[0069]

[0070] The unlabeled model image is defined as:

[0071]

[0072] The training model formula is:

[0073]

[0074] Among them, L is the training model, in the formula:

[0075]

[0076] It can be understood as the loss of the labeled model image, corresponding to the formula:

[0077]

[0078] It can be understood as the loss of unlabeled model images. The specific algorithm of each item in the formula will be described in detail in the subsequent text.

[0079] S104, receiving an image to be recognized;

[0080] This step receives an image to be identified that requires target detection.

[0081] S105: Detect the image to be identified using the training model to obtain a detection result of the identified object.

[0082] This step is to classify and identify the image to be identified by training the model. After this step, labels can be added to the classified and identified objects for annotation.

[0083] In this step, a target detection method is provided, which uses the pre-built student network and teacher network to train according to the average teacher strategy to obtain the labeled model image loss and the unlabeled model image loss. This method conforms to the semi-supervised learning method and can achieve adaptive labeled training results. The training model calculated by the labeled model image loss and the unlabeled model image loss can greatly improve the prediction accuracy of the received image to be recognized through the training model.

[0084] Figure 3 Schematic diagram of a target detection method process 300 in one embodiment of the present application. Figure 4 FIG. 4 is a schematic diagram of a target detection method process 400 in one embodiment of the present application. Figure 3 and Figure 4As shown, in one embodiment, the training of the labeled model image and the unlabeled model image through the pre-built student network and the teacher network according to the average teacher strategy to obtain the labeled model image loss value and the unlabeled model image loss value includes:

[0085] S201, inputting the labeled model image into the student network training to obtain the labeled model image loss;

[0086] This step provides a specific implementation method for training a student network to obtain the loss of the labeled model image.

[0087] S202, performing two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image;

[0088] This step is to enable the annotated model image to be capable of subsequent recognition processing when the scene is switched. The scene switching here can be understood as the concept of domain by those skilled in the art. The domain in the unlabeled model image is offset by noise interference.

[0089] The following formula explains the subsequent steps:

[0090] The first noise interference image is defined as:

[0091] The second noise interference image is defined as:

[0092] S203, inputting the first noise interference image into a teacher network training to obtain a first similarity feature and a candidate region of the teacher network;

[0093] In this step, the first similarity feature obtained by teacher network training is Figure 4 In The candidate region, i.e. Figure 4 In In addition, teacher network training can be understood as teacher network.

[0094] S204, inputting the candidate region and the second noise interference image into the student network for training to obtain a second similarity feature;

[0095] In this step, the student network is trained to obtain the second similarity feature, namely Figure 4 In

[0096] S205: Obtain an unlabeled model image loss by training according to the first similarity feature and the second similarity feature.

[0097] The main step is to establish the connection between the first similarity feature and the second similarity feature, and use it as the unlabeled model image loss for subsequent calculations.

[0098] In this embodiment, a labeled model image and an unlabeled model image are first received, and then trained separately through the student network and the teacher network respectively. The unlabeled model image pair is randomly enlarged (i.e., the interference in the above step) to obtain two different unlabeled model images, which are put into different neural networks for training. At the same time, one of the unlabeled model images trained in the teacher network will train the candidate area and the first. The candidate area is trained together with one of the unlabeled model images assigned to the student network to obtain the second similarity feature. The teacher network is trained by the unlabeled model image assigned to it to obtain the first similarity feature. A relationship is established between the first similarity feature and the second similarity feature, and then the established relationship is used as the unlabeled model image loss for subsequent calculations.

[0099] In order to better illustrate this embodiment, the following formula is used to express the implementation method of the above steps:

[0100] For two unlabeled model images:

[0101] Student network, using Will Parameterize and get the prediction function of the student network

[0102] Teacher Network, utilizing Will Parameterize and get the prediction function of the teacher network

[0103] From the above description we can get:

[0104]

[0105] This embodiment can ensure the accuracy of recognition when switching domains. The reason is that the physical quantity lost in the unlabeled model image can actually include three elements, namely the consistency problem mentioned above, the consistency between regions and the consistency between images, and the consistency within the image.

[0106] In order to provide a more specific description of the above consistency, we still use the formula to explain it in detail:

[0107] First, define the labeled model image for Linear regression annotation of ;

[0108] Define unlabeled model image

[0109] The final prediction annotation accuracy after domain shift depends on and

[0110] At the same time, for x t After randomly amplifying the interference, we get and

[0111] The candidate region in the above step is

[0112] The first similarity feature and the second similarity feature are and So It can be understood as the loss of unlabeled model images, and its essence is a mapping with a co-construction relationship.

[0113] for Further, you can to build.

[0114] is the detection result of all candidate regions in the student network and the teacher network. The expression is The matrix of .

[0115] like Figure 4 As shown, in one embodiment, the student network and the teacher network are Faster-R-CNN networks;

[0116] The Faster-R-CNN network has a Base CNN network, an RCNN network, and an RPN network;

[0117] The Base CNN network is used to extract image features, the RCNN network is used to extract corresponding similarity features based on the image features, and the RPN network is used to extract candidate regions of the image.

[0118] This embodiment provides a specific implementation of a student network and a teacher network, that is, both use a Faster-R-CNN network. In addition, the specific configuration of the Faster-R-CNN network is also disclosed.

[0119] like Figure 4 As shown, the configuration of this embodiment can be further explained by formula:

[0120] For input , the image features output by the Base CNN network are ,in, It can be understood as the output function of the Base CNN network;

[0121] The output function of the RPN network is , according to the previous configuration, we can get:

[0122]

[0123] The ROI pooling layer can extract Extract fixed vector ;

[0124] In addition and .

[0125] For the student network, the network architecture is the same as the Faster-R-CNN network architecture mentioned above, except that the student network will get a , so we get:

[0126]

[0127] Similarly, there are and

[0128] Define a calculation coefficient,

[0129] Among them, C is the data configuration of the preset classification data set, is the predicted probability of the jth preset classification data set.

[0130] The relationship between domains can be expressed as:

[0131]

[0132] The above method is the calculation of domain level consistency.

[0133] The calculation method for consistency between graphs is as follows:

[0134] Define two candidate regions, denoted as

[0135] In summary, we can get:

[0136]

[0137] The consistency between graphs can be expressed as:

[0138]

[0139] Finally, let’s discuss the problem of intra-graph consistency in the student network:

[0140]

[0141] The matrix:

[0142] Using the above formula, the final expression of the consistency in the graph is:

[0143]

[0144] The original formula can be:

[0145]

[0146] Further expanded to:

[0147]

[0148] A training model is constructed based on the inter-domain consistency, inter-image consistency and intra-image consistency in the formula, as well as the image loss of the annotation model.

[0149] Figure 5 This is a schematic diagram of a target detection method process 500 in one embodiment of the present application. Figure 6 FIG. 1 is a schematic diagram showing the comparison of target detection accuracy after the weight coefficient is changed in one embodiment of the present application. Figure 5 and Figure 6 As shown, in one embodiment, after the step of calculating the training model according to the loss value of the labeled model image and the loss value of the unlabeled model image, the method further includes:

[0150] S301, adjusting the weight of the unlabeled model image loss;

[0151] In this step;

[0152] According to the following formula in the previous article:

[0153]

[0154] As long as the value of the weight coefficient λ is adjusted, the weight coefficient of the unlabeled model image loss can be adjusted.

[0155] S302: Obtain a training model according to the labeled model image loss and the adjusted unlabeled model image loss.

[0156] In this step, the unlabeled model image loss with adjusted weights is calculated into the training model, and the subsequent target detection step is performed.

[0157] This step provides a method for adjusting the unlabeled model image loss. In the application, the unlabeled model image loss can affect the final recognition accuracy in the training model. Therefore, the final recognition result can be adjusted by adjusting the unlabeled model image loss. Figure 6 It can also be seen that the change of weight coefficient affects the final result of target detection.

[0158] Figure 7 This is a schematic diagram of comparing the target detection accuracy after the smoothing coefficient is changed in one embodiment of the present application; Figure 8 Schematic diagram of the comparison of target detection accuracy of different methods after changing the weight coefficient and smoothing coefficient.

[0159] like Figure 7 and Figure 8 As shown, in one embodiment, inputting the labeled model image into the student network training to obtain the labeled model image loss includes:

[0160] The labeled model image is input into the student network and trained using a stochastic gradient descent algorithm to obtain a labeled model image.

[0161] In this embodiment, it is easier to understand with the following formula (the derivation process is given above and will not be repeated here), where α is the smoothing coefficient.

[0162]

[0163] Experimental description:

[0164]

[0165] Table 1 above is a performance comparison of different methods on the foggy urban scene verification set. The vertical axis shows the different methods used, and the horizontal axis sets person, rider, car, bus, train, bicycle, etc. The corresponding data below is the recognition accuracy.

[0166] Source-only in the vertical direction is a method in the prior art. This method simply trains the labeled model image through CNN, and recognizes the image through the trained model. The following DA is restructured using 50 layers of ResNet. The methods below DA are all methods protected by this application, but the configurations are different. From Table 1, it can be seen that the accuracy of the method of this application at all levels is better than the simple training of the prior art and the DA method. In addition, Fig. 9 This is a schematic diagram of an experimental comparison of target detection under different methods. Fig. 9 The target detection results of different methods can also be seen.

[0167]

[0168] Table 2 above shows the average accuracy of different methods for cars in the city.

[0169] Table 3.The mean Average Precision(mAP) of different models onSyn2Real detection dataset for S→O / Y transfers.

[0170]

[0171] Table 3 above shows the average accuracy of other configurations of different methods

[0172] Fig.10 This is another experimental comparison diagram of target detection accuracy under different methods. Fig.10 As shown in Figure 2, different methods have a great impact on the recognition results.

[0173] Fig.11 This is a schematic diagram of the target detection device architecture in one embodiment of the present application. Fig.11 As shown, the present application also provides a target detection device, which includes:

[0174] The receiving module 101 is used to receive a labeled model image and an unlabeled model image, wherein the labeled model image is an image with a pre-built recognition object classification data set, and the unlabeled model image is an image without a pre-built recognition object classification data set;

[0175] A training module 102 is used to train the labeled model image and the unlabeled model image through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss;

[0176] A calculation module 103, configured to calculate a training model according to the labeled model image loss and the unlabeled model image loss;

[0177] The receiving module 101 is further used to receive an image to be identified;

[0178] The recognition module 104 is used to detect the image to be recognized using the training model to obtain a detection result of the recognized object.

[0179] In one embodiment, the apparatus comprises:

[0180] The training module 102 is further used to input the labeled model image into the student network for training to obtain the labeled model image loss;

[0181] A noise module 105, configured to perform two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image;

[0182] The training module 102 is further used to perform two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image; is further used to input the candidate area and the second noise interference image into the student network for training to obtain a second similarity feature; and is further used to obtain the unlabeled model image loss based on the first similarity feature and the second similarity feature.

[0183] In one embodiment, the apparatus comprises:

[0184] An adjustment module 106, configured to adjust the weight of the unlabeled model image loss;

[0185] The calculation module 103 is further used to calculate a training model according to the labeled model image loss and the adjusted unlabeled model image loss.

[0186] The present application also provides a target detection device, the device comprising: a processor and a memory;

[0187] The memory stores an application program executable by the processor, which is used to enable the processor to execute the steps of the target detection method.

[0188] In one embodiment, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the target detection method steps when the program is executed by a processor.

[0189] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target detection method, characterized in that: The method includes: Receiving a labeled model image and an unlabeled model image, wherein the labeled model image is an image with a pre-built recognition object classification data set, and the unlabeled model image is an image without a pre-built recognition object classification data set; The labeled model image and the unlabeled model image are trained according to the average teacher strategy through the pre-built student network and the teacher network to obtain the labeled model image loss and the unlabeled model image loss; wherein the student network is used to process the labeled model image, and the teacher network is used to process the unlabeled model image, and the teacher network will guide the student network to achieve the effect of having common attributes; the labeled model image and the unlabeled model image are trained according to the average teacher strategy through the pre-built student network and the teacher network to obtain the labeled model image loss and the unlabeled model image loss, including: inputting the labeled model image into the student network training to obtain the labeled model image loss; performing two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image; inputting the first noise interference image into the teacher network training to obtain a first similarity feature and a candidate area of ​​the teacher network; inputting the candidate area and the second noise interference image into the student network training to obtain a second similarity feature; calculating the unlabeled model image loss according to the first similarity feature and the second similarity feature; A training model is obtained by calculating the loss of the labeled model image and the loss of the unlabeled model image; Receiving an image to be recognized; The training model is used to detect the image to be identified to obtain a detection result of the identified object.

2. The target detection method according to claim 1, characterized in that: The student network and the teacher network are Faster-R-CNN networks; The Faster-R-CNN network has a Base CNN network, an RCNN network, and an RPN network; The Base CNN network is used to extract image features, the RCNN network is used to extract corresponding visual features based on the image features, and the RPN network is used to extract candidate regions of the image.

3. The target detection method according to claim 1, characterized in that: After the step of calculating the training model according to the labeled model image loss and the unlabeled model image loss, the method further includes: Adjusting the weight of the unlabeled model image loss; A training model is obtained by calculating the loss of the labeled model image and the adjusted loss of the unlabeled model image.

4. The target detection method according to claim 1, characterized in that: The step of inputting the labeled model image into the student network training to obtain the labeled model image loss comprises: The labeled model image is input into the student network and trained using a stochastic gradient descent algorithm to obtain a labeled model image.

5. A target detection device, characterized in that: The device includes: A receiving module, used for receiving annotated model images and unlabeled model images, wherein the annotated model images are image sets with pre-built recognition object classification data sets, and the unlabeled model images are image sets without pre-built recognition object classification data sets; A training module, used to train the labeled model image and the unlabeled model image through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss; wherein the student network is used to process the labeled model image, and the teacher network is used to process the unlabeled model image, and the teacher network will guide the student network to achieve the effect of having common attributes; the labeled model image and the unlabeled model image are trained through a pre-built student network and a teacher network according to an average teacher strategy to obtain a labeled model image loss and an unlabeled model image loss, including: inputting the labeled model image into the student network for training to obtain a labeled model image loss; performing two different noise interferences on the unlabeled model image to obtain a first noise interference image and a second noise interference image; inputting the first noise interference image into the teacher network for training to obtain a first similarity feature and a candidate area of ​​the teacher network; inputting the candidate area and the second noise interference image into the student network for training to obtain a second similarity feature; and calculating the unlabeled model image loss according to the first similarity feature and the second similarity feature; A calculation module, used for calculating a training model according to the labeled model image loss and the unlabeled model image loss; The receiving module is further used to receive the image to be identified; The recognition module is used to detect the image to be recognized using the training model to obtain a detection result of the recognized object.

6. The target detection device according to claim 5, characterized in that: The device includes: An adjustment module, used for adjusting the weight of the unlabeled model image loss; The calculation module is also used to calculate the training model according to the labeled model image loss and the adjusted unlabeled model image loss.

7. A target detection device, characterized in that: The device comprises: a processor and a memory; The memory stores an application program executable by the processor, which is used to enable the processor to perform the steps of the target detection method according to any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the target detection method described in any one of claims 1 to 4 are implemented.

Images