Threshold adaptive object detection network training method and device

Abstract

The application provides a threshold adaptive object detection network training method and device. The method inputs a training set containing a target to be identified into an object detection network to train backbone weight parameters of the object detection network. After the backbone weight parameters reach an expected value, a threshold convolution branch is added to the object detection network to provide different confidence levels for different objects. The training set containing the target to be identified and a classification label of the target to be identified is input into the object detection network, and the backbone parameters are locked to train branch weight parameters to an expected value to obtain an object detection network capable of adaptively generating confidence levels according to object categories. Through adaptive confidence generation, the generalization of the object detection network can be improved, and the false detection rate can be reduced.

Description

Technical Field

[0001] This application relates to the field of object detection technology, and in particular to a threshold-adaptive object detection network training method and apparatus. Background Technology

[0002] When identifying objects using an object detection network, different confidence thresholds need to be set according to different detection categories. This allows the network to filter the identification results and determine whether the results are positive samples.

[0003] When building an object detection network, a confidence threshold can be set based on the experience of technicians. However, setting a confidence threshold based on experience requires high-quality data of the target object. With a high confidence threshold set based on experience, false negatives are likely to occur when the target object is some distance from the target class center but still belongs to the target class. With a low confidence threshold set based on experience, false positives are likely to occur for similar targets.

[0004] Furthermore, based on the detection results of the object detection network on the test set, curves of detection rate and false positive rate can be obtained, and confidence thresholds can be set according to statistical indicators. However, in practical applications, when the actual application scenario differs significantly from the test set scenario, this method is prone to missed detections. Missed detections are also likely to occur when the data quality is low. Summary of the Invention

[0005] This application provides a threshold-adaptive object detection network training method and apparatus to solve the problem that the confidence threshold set according to the scene of the test set is inaccurate and easily leads to missed detections.

[0006] Firstly, this application provides a threshold-adaptive object detection network training method, including:

[0007] The training set containing the target to be identified is input into the object detection network to train the backbone weight parameters of the object detection network.

[0008] A threshold convolution branch is added to the object detection network; the threshold convolution branch is used to provide a confidence threshold for the recognition result output by the object detection network; the confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample.

[0009] The training set containing the target to be identified and its classification label is input into the object detection network to train the branch weight parameters of the threshold convolutional branch; and the backbone weight parameters are locked.

[0010] If the branch weight parameters are detected to have reached the expected value, the current object detection network and its backbone weight parameters and branch weight parameters are output.

[0011] In some feasible embodiments, the recognition results output by the object detection network include the class confidence score, bounding box regression state, and bounding box confidence score of the target to be identified, and further include:

[0012] The loss value of class confidence is calculated based on the loss function of class confidence, and the loss value of box regression state is calculated based on the loss function of setting box regression state, and the loss value of box confidence is calculated based on the loss function of box confidence.

[0013] If the loss value of the class confidence is greater than the class confidence loss threshold, and / or the loss value of the box regression state is greater than the box regression state loss threshold, and / or the loss value of the box confidence is greater than the box confidence loss threshold, then the backbone weight parameters of the object detection network are modified according to the loss value.

[0014] If the loss value of the class confidence is less than or equal to the class confidence loss threshold, and the loss value of the box regression state is less than or equal to the box regression state loss threshold, and the loss value of the box confidence is less than or equal to the box confidence loss threshold, then the step of adding a threshold convolution branch to the object detection network is executed.

[0015] In some feasible embodiments, when adding a threshold convolutional branch to the object detection network, the method further includes:

[0016] A threshold loss function is set for the added threshold convolution branch, and the threshold loss function is set based on the correct detection rate of the target to be identified in a single training process;

[0017] The threshold loss function is used to calculate the threshold loss value of the object detection network with the added threshold convolution branch;

[0018] If the threshold loss value is greater than the threshold loss value, then the branch weight parameters of the object detection network are modified according to the threshold loss value;

[0019] If the threshold loss value is less than or equal to the threshold loss threshold, then output the current object detection network including the threshold convolution branch.

[0020] In some feasible embodiments, the threshold-adaptive object detection network training method further includes:

[0021] Set the average confidence threshold;

[0022] Based on the average confidence threshold, an L2 constraint associated with the confidence threshold is added to the bounded loss function.

[0023] Secondly, this application provides an object detection method, including:

[0024] The test image containing the target to be identified is input into an object detection network including a threshold convolution branch. The object detection network is trained according to the threshold adaptive object detection network training method described in the training method embodiment.

[0025] Based on the object detection network including the threshold convolution branch, the confidence score of the target to be identified is calculated according to the center coordinates of the target to be identified, and the confidence score threshold of the target to be identified is searched.

[0026] If the confidence level of the target to be identified is greater than or equal to the confidence level threshold, the identification result associated with the current confidence level is output.

[0027] In some feasible embodiments, the number of recognition results for the target to be identified is at least 1, and the output of the object detection network includes a non-maximum suppression module, which is used to output the recognition result with the highest confidence.

[0028] Thirdly, this application provides a threshold-adaptive object detection network training device, comprising: a first training module, a control module, and a second training module;

[0029] The first training module is used to input a training set containing the target to be identified into the object detection network to train the backbone weight parameters of the object detection network.

[0030] The control module is used to add a threshold convolution branch to the object detection network; the threshold convolution branch is used to provide a confidence threshold for the recognition result output by the object detection network; the confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample.

[0031] The second training module is used to input a training set containing the target to be identified and its classification label into the object detection network to train the branch weight parameters of the threshold convolution branch; and to lock the backbone weight parameters.

[0032] When the control module detects that the branch weight parameter has reached the expected value, it outputs the current object detection network and the backbone weight parameter and branch weight parameter of the current object detection network.

[0033] Fourthly, this application provides an object detection device, including: a data acquisition module, a computing module, and a decision-making module;

[0034] The acquisition module is used to input the test image containing the target to be identified into an object detection network including a threshold convolution branch;

[0035] The computation module is used to calculate the confidence score of the target to be identified based on the center coordinates of the target to be identified, and to search for the confidence score threshold of the target to be identified, based on the object detection network including the threshold convolution branch.

[0036] The decision module is used to output an identification result associated with the current confidence level when the confidence level of the target to be identified is greater than or equal to the confidence level threshold.

[0037] Fifthly, this application provides an electronic device, characterized in that it comprises:

[0038] A processor, and a memory communicatively connected to the processor;

[0039] The memory stores instructions that can be executed by the processor to cause the processor to perform the steps of the training method described in the first aspect.

[0040] Sixthly, this application provides a computer-readable storage medium, comprising:

[0041] The computer-readable storage medium includes at least one computer instruction for causing the computer to perform the steps of the training method as described in the first aspect.

[0042] As described above, this application provides a threshold-adaptive object detection network training method and apparatus. The method inputs a training set containing the target to be identified into the object detection network to train the network's backbone weight parameters. After the backbone weight parameters reach their desired values, a threshold convolutional branch is added to the object detection network to provide different confidence levels for different objects. The training set containing the target to be identified and its classification label is input into the object detection network, and the backbone parameters are locked to train the branch weight parameters to their desired values, resulting in an object detection network that can adaptively generate confidence levels based on object categories. By adaptively generating confidence levels, the generalization ability of the object detection network can be improved, and the false detection rate can be reduced. Attached Figure Description

[0043] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A schematic diagram illustrating the training method of the threshold adaptive object detection network provided in the embodiments of this application;

[0045] Figure 2A schematic diagram of the backbone weight parameters training of the object detection network provided in an embodiment of this application;

[0046] Figure 3 A schematic diagram illustrating the training of branch weight parameters for an object detection network provided in an embodiment of this application;

[0047] Figure 4 This is a schematic diagram of the object detection method provided in the embodiments of this application. Detailed Implementation

[0048] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0049] When identifying targets in a test image, the image can be input into an object detection network. The object detection network determines the validity of the target identification result based on a confidence threshold. However, various factors such as the target's pose, sharpness, and size in the test image can affect the confidence calculation of the object detection network, leading to some bias in the calculated confidence results. Therefore, when using a uniform global confidence threshold, the confidence threshold for some targets may not be suitable for the application scenario, resulting in missed detections or false positives.

[0050] like Figure 1 As shown, this application provides a threshold-adaptive object detection network training method to solve the above problems, the method comprising:

[0051] S100: Input the training set containing the target to be identified into the object detection network to train the backbone weight parameters of the object detection network.

[0052] In some embodiments, the object detection network can be a YOLOX (You Only Look Once, object detection model) network. The object detection network may include a Feature Fusion Network (FPN) and a Prediction Output Network (Header). The training set includes various targets to be identified, and the scenes, sizes, and poses of the targets can be different states, in order to improve the generalization ability of the model through multiple training samples, thereby improving the recognition accuracy.

[0053] When outputting recognition results, object detection networks can output class confidence, bounding box regression status, and bounding box confidence. The backbone weight parameters can be the weight parameters of the convolutional branches containing the class confidence, bounding box regression status, and bounding box confidence. By training the backbone weight parameters, the accuracy of the class confidence, bounding box regression status, and bounding box confidence can be improved, thereby improving the accuracy of the object recognition results.

[0054] like Figure 2 As shown, in some embodiments, the recognition results output by the object detection network include the class confidence, bounding box regression state, and bounding box confidence of the target to be identified. The training method further includes:

[0055] The loss value of class confidence is calculated based on the loss function of class confidence, and the loss value of box regression state is calculated based on the loss function of setting box regression state, and the loss value of box confidence is calculated based on the loss function of box confidence.

[0056] If the loss value of the class confidence is greater than the class confidence loss threshold, and / or the loss value of the box regression state is greater than the box regression state loss threshold, and / or the loss value of the box confidence is greater than the box confidence loss threshold, then the backbone weight parameters of the object detection network are modified according to the loss value.

[0057] If the loss value of the class confidence is less than or equal to the class confidence loss threshold, and the loss value of the box regression state is less than or equal to the box regression state loss threshold, and the loss value of the box confidence is less than or equal to the box confidence loss threshold, then the step of adding a threshold convolution branch to the object detection network is executed.

[0058] Class confidence, bounding box regression state, and bounding box confidence can be used to determine whether the recognition results are effective. Therefore, when training an object detection network using a training set, class confidence loss function, bounding box regression state loss function, and bounding box confidence loss function can be set based on class confidence, bounding box regression state, and bounding box confidence. The loss value of each convolutional branch can be calculated by setting the loss function, and then the training effect can be judged by the loss value.

[0059] When judging the training effect using the loss value, the backbone weight parameters of each convolutional branch need to reach the expected value to ensure the accuracy of the model's recognition results. Therefore, during training, the backbone weight parameters are considered to have been trained to the optimal level only when all the loss values ​​calculated by the loss function are less than the loss threshold. After the backbone weights have been trained to the optimal level, the step of adding threshold convolutional branches to the object detection network can be performed to train the branch weight parameters of the threshold convolutional branches.

[0060] S200: Add a threshold convolution branch to the object detection network;

[0061] like Figure 1 As shown, a threshold convolution branch can be added to the object detection network. This threshold convolution branch provides a confidence threshold for the recognition result output by the object detection network. The confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample. The imaging quality of each object will vary; various factors such as the scene, size, and pose of the object in the image can affect the object recognition result. Therefore, by setting a threshold convolution branch in the object detection network, a confidence threshold can be provided for each type of object, thereby improving the generalization of the object detection network and making it suitable for more object recognition. For example, if the targets to be recognized include electric vehicles, motorcycles, and bicycles, the object detection network including the threshold convolution branch will configure different confidence levels for the recognition results of electric vehicles, motorcycles, and bicycles to determine whether the recognition result is a valid result.

[0062] like Figure 3 As shown, in some embodiments, when adding a threshold convolutional branch to the object detection network, the following further methods are included:

[0063] A threshold loss function is set for the added threshold convolution branch, and the threshold loss function is set based on the correct detection rate of the target to be identified in a single training process;

[0064] The threshold loss function is used to calculate the threshold loss value of the object detection network with the added threshold convolution branch;

[0065] If the threshold loss value is greater than the threshold loss value, then the branch weight parameters of the object detection network are modified according to the threshold loss value;

[0066] If the threshold loss value is less than or equal to the threshold loss threshold, then output the current object detection network including the threshold convolution branch.

[0067] The bounded loss function differs from the class confidence loss function, box regression state loss function, and box confidence loss function in the above embodiments. The bounded loss function is jointly determined by class confidence, box regression state, box confidence, and confidence threshold. Therefore, the bounded loss function uses the correct detection rate in a single training iteration as the training objective; that is, in some embodiments, the correct detection rate of all boxes in a batch is used as the training objective.

[0068] The bounded loss value during training can be calculated based on the bounded loss function and the correct detection rate. Correspondingly, to ensure training effectiveness and efficiency, a bounded loss threshold can be set to judge the training performance. When the bounded loss value exceeds the bounded loss threshold, the branch weight parameters need to be modified according to the loss value. It should be noted that when modifying the branch weight parameters, the branch weight parameters of the object detection network can be actively modified, or the object detection network can iteratively train the branch weight parameters based on the backpropagation of the bounded loss value within the network to modify the branch weight parameters.

[0069] In some embodiments, when calculating the bounded loss value using the bounded loss function, an L2 constraint may also be added to the confidence threshold. The steps for adding the L2 constraint include:

[0070] Set the average confidence threshold;

[0071] Based on the average confidence threshold, an L2 constraint associated with the confidence threshold is added to the bounded loss function.

[0072] In the above embodiments, the following formula can be used:

[0073] Loss_thd_L2=1 / n*Σ(Thd(i)–Thd_average)^2;

[0074] Where Thd(i) is the confidence threshold value corresponding to the center of each object; thd_average is the average confidence threshold, which can be 0.5 during training.

[0075] L2 constraints balance the relatively sparse, spiked weight vectors, improving the utilization of input data during computation and making the calculation results more uniform. This allows for better training results even with training samples featuring unique characteristics, thereby enhancing the generalization ability of object detection networks and ultimately improving object recognition accuracy.

[0076] S300: Input the training set containing the target to be identified and the target classification label into the object detection network to train the branch weight parameters of the threshold convolution branch; and lock the backbone weight parameters;

[0077] After adding a threshold convolutional branch to the object detection network, a new training set can be deployed to train the network. The new training set includes the target to be identified and its classification label, which helps the threshold convolutional branch learn based on the target and its classification label during training, thereby providing an appropriate confidence threshold for each class of target to be identified.

[0078] Understandably, the training samples in the new training set should be more diverse in form. For example, the targets to be identified in the training set can include features such as imaging degradation, small targets, and multiple poses, so as to learn the threshold convolution branch in various scenarios. This makes the threshold convolution branch applicable to more application scenarios, and thus, when facing each target to be identified, it can enable the target to achieve optimal classification at its respective confidence threshold.

[0079] Locking the backbone weights avoids further training on backbone weights that have already reached a good level, thus preventing increased training time and improving model training efficiency. Training only the branch weights using a bounded loss function allows the gated convolutional branches to iteratively learn based on the bounded loss value and the bounded loss threshold.

[0080] S400: If the branch weight parameter is detected to have reached the expected value, the current object detection network and its backbone weight parameter and branch weight parameter are output.

[0081] When the branch weight parameters reach the expected value, i.e., the bounded loss calculated according to the bounded loss function is lower than the bounded loss threshold, it indicates that the current branch weight parameters have reached the optimal state. At this point, the current object detection network, as well as the backbone weight parameters and branch weight parameters of the current object detection network, can be output.

[0082] Current object detection networks employ a two-step training approach, training the backbone weights and branch weights separately. During object detection, a threshold convolution branch outputs a confidence threshold corresponding to the target. When training the branch weights, training samples of the target with the desired classification labels are used, and the scenarios of the training samples are enriched to improve the generalization ability of the threshold convolution branch. This allows the threshold convolution branch to output a confidence threshold corresponding to the target even when facing diverse application scenarios and target states, thereby improving the accuracy of the object detection network and reducing the probability of false positives and false negatives.

[0083] like Figure 4 As shown, in some embodiments, this application provides an object detection method, the detection method comprising:

[0084] S700: Input the test image containing the target to be identified into an object detection network including a threshold convolution branch. The object detection network is trained according to the threshold adaptive object detection network training method described in the training method class embodiment.

[0085] S800: Based on the object detection network including the threshold convolution branch, calculate the confidence of the target to be identified according to the center coordinates of the target to be identified, and search the confidence threshold of the target to be identified;

[0086] S900: If the confidence level of the target to be identified is greater than or equal to the confidence level threshold, then output the identification result associated with the current confidence level.

[0087] In the described embodiment, when the image to be identified is input into the object detection network for recognition, the object detection network directly obtains the confidence threshold of the target to be identified in the image. For example, if the target to be identified is an electric vehicle, the confidence threshold for an electric vehicle obtained during training is 0.8. The object detection network directly outputs the confidence threshold of 0.8 for the electric vehicle to determine whether the recognition result is valid. When the confidence scores in the recognition results for the electric vehicle are all not lower than 0.8, the recognition result for the electric vehicle can be determined to be valid, and the recognition result for the target to be identified is "electric vehicle".

[0088] The image to be tested is input into an object detection network that includes a threshold convolution branch. The object detection network can provide a confidence threshold associated with the target to be identified. The branch weight parameters of the threshold convolution branch in the object detection network are obtained by training and optimizing the target to be identified in various scenarios. Therefore, it is beneficial to improve the adaptability of the confidence threshold to the target to be identified, thereby improving the recognition accuracy of the object detection network.

[0089] In some embodiments, the number of recognition results for the target to be identified is at least 1, and the output of the object detection network includes a non-maximum suppression module, which is used to output the recognition result with the highest confidence.

[0090] Some targets to be identified are highly similar to other targets, such as electric vehicles and motorcycles. Therefore, object detection networks will generate multiple identification results during the identification process. When multiple identification results are generated, by setting a non-maximum suppression module, the identification results can be filtered according to their confidence level, and the identification result with the highest confidence level can be output.

[0091] Understandably, if the confidence level of the recognition result is lower than the confidence threshold, it is directly determined as an invalid recognition result. The non-maximum suppression module only filters recognition results that meet the confidence threshold value to improve object detection efficiency.

[0092] In some embodiments, this application provides a threshold-adaptive object detection network training device, including: a first training module, a control module, and a second training module;

[0093] The first training module is used to input a training set containing the target to be identified into the object detection network to train the backbone weight parameters of the object detection network.

[0094] The control module is used to add a threshold convolution branch to the object detection network; the threshold convolution branch is used to provide a confidence threshold for the recognition result output by the object detection network; the confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample.

[0095] The second training module is used to input a training set containing the target to be identified and its classification label into the object detection network to train the branch weight parameters of the threshold convolution branch; and to lock the backbone weight parameters.

[0096] When the control module detects that the branch weight parameter has reached the expected value, it outputs the current object detection network and the backbone weight parameter and branch weight parameter of the current object detection network.

[0097] In some embodiments, this application provides an object detection device, including: a data acquisition module, a computing module, and a decision-making module;

[0098] The acquisition module is used to input the test image containing the target to be identified into an object detection network including a threshold convolution branch;

[0099] The computation module is used to calculate the confidence score of the target to be identified based on the center coordinates of the target to be identified, and to search for the confidence score threshold of the target to be identified, based on the object detection network including the threshold convolution branch.

[0100] The decision module is used to output an identification result associated with the current confidence level when the confidence level of the target to be identified is greater than or equal to the confidence level threshold.

[0101] In some embodiments, this application provides an electronic device, including:

[0102] A processor, and a memory communicatively connected to the processor;

[0103] The memory stores instructions that can be executed by the processor to cause the processor to perform the steps of the method described in the training method embodiment.

[0104] In some embodiments, this application provides a computer-readable storage medium, comprising:

[0105] The computer-readable storage medium includes at least one computer instruction for causing the computer to perform steps of the method described in the training method-like embodiments.

[0106] As described above, this application provides a threshold-adaptive object detection network training method and apparatus. The method inputs a training set containing the target to be identified into the object detection network to train the network's backbone weight parameters. After the backbone weight parameters reach their desired values, a threshold convolutional branch is added to the object detection network to provide different confidence levels for different objects. The training set containing the target to be identified and its classification label is input into the object detection network, and the backbone parameters are locked to train the branch weight parameters to their desired values, resulting in an object detection network that can adaptively generate confidence levels based on object categories. By adaptively generating confidence levels, the generalization ability of the object detection network can be improved, and the false detection rate can be reduced.

[0107] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A threshold-adaptive object detection network training method, characterized in that, include: The training set containing the target to be identified is input into the object detection network to train the backbone weight parameters of the object detection network. A threshold convolution branch is added to the object detection network; the threshold convolution branch is used to provide a confidence threshold for the recognition result output by the object detection network; the confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample. The training set containing the target to be identified and its classification label is input into the object detection network to train the branch weight parameters of the threshold convolutional branch; and the backbone weight parameters are locked. If the branch weight parameter is detected to have reached the expected value, the current object detection network and its backbone weight parameter and branch weight parameter are output.

2. The threshold-adaptive object detection network training method according to claim 1, wherein the recognition result output by the object detection network includes the class confidence, bounding box regression state, and bounding box confidence of the target to be identified, characterized in that, Also includes: The loss value of class confidence is calculated based on the loss function of class confidence, and the loss value of box regression state is calculated based on the loss function of setting box regression state, and the loss value of box confidence is calculated based on the loss function of box confidence. If the loss value of the class confidence is greater than the class confidence loss threshold, and / or, the loss value of the box regression state is greater than the box regression state loss threshold, and / or, the loss value of the box confidence is greater than the box confidence loss threshold, then the backbone weight parameters of the object detection network are modified according to the loss value. If the loss value of the class confidence is less than or equal to the class confidence loss threshold, and the loss value of the box regression state is less than or equal to the box regression state loss threshold, and the loss value of the box confidence is less than or equal to the box confidence loss threshold, then the step of adding a threshold convolution branch to the object detection network is executed.

3. The threshold-adaptive object detection network training method according to claim 2, characterized in that, When adding a threshold convolution branch to the object detection network, the following is also included: A threshold loss function is set for the added threshold convolution branch, and the threshold loss function is set based on the correct detection rate of the target to be identified in a single training process; The threshold loss function is used to calculate the threshold loss value of the object detection network with the added threshold convolution branch; If the threshold loss value is greater than the threshold loss value, then the branch weight parameters of the object detection network are modified according to the threshold loss value; If the threshold loss value is less than or equal to the threshold loss threshold, then output the current object detection network including the threshold convolution branch.

4. The threshold-adaptive object detection network training method according to claim 3, characterized in that, Also includes: Set the average confidence threshold; Based on the average confidence threshold, an L2 constraint associated with the confidence threshold is added to the bounded loss function.

5. An object detection method, characterized in that, include: The test image containing the target to be identified is input into an object detection network including a threshold convolution branch, which is trained according to the threshold adaptive object detection network training method of any one of claims 1-4. Based on the object detection network including the threshold convolution branch, the confidence level of the target to be identified is calculated according to the center coordinates of the target to be identified, and the confidence level threshold of the target to be identified is obtained. If the confidence level of the target to be identified is greater than or equal to the confidence level threshold, the identification result associated with the current confidence level is output.

6. The method according to claim 5, characterized in that, The number of recognition results for the target to be identified is at least 1. The output of the object detection network includes a non-maximum suppression module, which is used to output the recognition result with the highest confidence.

7. A threshold-adaptive object detection network training device, characterized in that, include: First training module, control module, second training module; The first training module is used to input a training set containing the target to be identified into the object detection network to train the backbone weight parameters of the object detection network. The control module is used to add a threshold convolution branch to the object detection network; The threshold convolutional branch is used to provide a confidence threshold for the recognition result output by the object detection network; the confidence threshold is used to determine whether the recognition result output by the object detection network is a positive sample; The second training module is used to input a training set containing the target to be identified and its classification label into the object detection network to train the branch weight parameters of the threshold convolution branch; and to lock the backbone weight parameters. When the control module detects that the branch weight parameter has reached the expected value, it outputs the current object detection network and the backbone weight parameter and branch weight parameter of the current object detection network.

8. An object detection device, characterized in that, include: Data acquisition module, computation module, decision-making module; The acquisition module is used to input the test image containing the target to be identified into an object detection network including a threshold convolution branch; the object detection network is trained according to the threshold adaptive object detection network training method according to any one of claims 1-4; The computation module is used to calculate the confidence level of the target to be identified based on the center coordinates of the target to be identified, and to obtain the confidence level threshold of the target to be identified, based on the object detection network including the threshold convolution branch. The decision module is used to output an identification result associated with the current confidence level when the confidence level of the target to be identified is greater than or equal to the confidence level threshold.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores instructions that are executed by the processor to cause the processor to perform the steps of the method according to any one of claims 1-4.

10. A computer-readable storage medium, characterized in that, include: The computer-readable storage medium includes at least one computer instruction for causing the computer to perform the steps of the method as described in any one of claims 1-4.

Images