CN109784349B - Image target detection model establishing method, device, storage medium and program product - Google Patents

Abstract

The present invention provides a method and device for establishing an image target detection model. The occlusion image samples are used to train the feature occlusion confrontation network model, and the occlusion mask of the image sample is obtained through the feature occlusion confrontation network model. In this way, the detection network model is performed. During the training, an occlusion mask is added to the feature map of the image samples used for training, and the occlusion mask is obtained by using the trained feature occlusion adversarial network model. In this way, since the feature occlusion adversarial network model is obtained by training occlusion image samples, a feature occlusion adversarial network model that generates better masks can be trained, and then the feature occlusion adversarial network is used to obtain an occlusion mask for training the detection network model. Therefore, the detection network model can be trained by using image samples with better occlusion masks, so that the detection network model can be fully trained for the occlusion situation, and the accuracy of the detection network model for the detection of occluded objects can be improved.

Description

Image target detection model establishment method, device, storage medium and program product

technical field

The invention relates to the field of artificial intelligence, in particular to a method and device for establishing an image target detection model, a storage medium and a program product.

Background technique

At present, detection algorithms based on deep convolutional neural networks have become the mainstream methods of image target detection, such as YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector) algorithms, which directly transform the problem of target frame positioning into regression. Problems are processed with faster detection speed. In the application of image target detection, the object to be detected in the image is occluded, and the current detection model seldom considers the occlusion situation and cannot achieve accurate detection of occluded objects.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a method and device for establishing an image target detection model, a storage medium and a program product, so as to improve the accuracy of the detection of occluded objects.

For achieving the above object, the present invention has the following technical solutions:

A model building method for image target detection, comprising:

occluding the first image sample to obtain a first occlusion image sample;

Use the first occlusion image sample to train a feature occlusion confrontation network model, where the feature occlusion confrontation network model is used to obtain an occlusion mask of the image sample based on the confrontation network;

A detection network model is trained, and an occlusion mask is added to the feature map of the second image sample used for training using the trained feature occlusion confrontation network model, and the detection network model is used for image target detection based on deep learning.

Optionally, the method further includes: using the feature map added with the occlusion mask as an image sample to continue the training of the feature occlusion confrontation network model.

Optionally, the blocking of the first image sample includes:

determining the candidate frame with the maximum positioning accuracy in the first image sample;

Map a sliding window of preset size to an image sample, and use background pixels to fill the image area where the sliding window is located;

The detection network model is used to detect the filled image samples, and the sliding window at the position where the candidate frame has the largest detection network loss is used as the occlusion position of the first image sample to obtain the first occlusion image sample.

Optionally, in the training of the detection network model, a non-maximum value suppression algorithm is used to determine the prediction frame, wherein the prediction frame evaluation index in the non-maximum value suppression algorithm is determined by the positioning accuracy and the position accuracy of different candidate frames. Location accuracy is determined.

Optionally, the calculation formula of the candidate frame evaluation index CL is:

CL=γ×sore _class +(1-γ)×score _location ; where γ is the hyperparameter, sore _class is the classification accuracy, and score _location is the positioning accuracy;

The formula of the non-maximum suppression algorithm is:

Among them, CL is the positioning accuracy of different candidate boxes.

Optionally, a predicted IoU network model is preset, and the predicted IoU network model is used to obtain the positioning accuracy of different candidate frames.

Optionally, the training method of the predicted IoU network model includes:

generating a set of candidate boxes for the third image sample;

Obtain the positioning accuracy of each candidate frame in the candidate frame set;

removing candidate frames whose positioning accuracy is less than a preset threshold in the candidate frame set to determine the training set of the third occlusion image sample;

Use the training set to train a predictive IoU network model.

Optionally, the detection network model includes a global pooling module and a pooling classification module, and the pooling classification module is used to classify the globally pooled feature map according to the corresponding position of the pooling mask to obtain a classification. vector.

A model building device for image target detection, comprising:

an occlusion sample obtaining unit, configured to occlude the first image sample to obtain the first occlusion image sample;

an adversarial network training unit, configured to use the first occlusion image sample to train a feature occlusion adversarial network model, where the feature occlusion adversarial network model is used to obtain an occlusion mask of the image sample based on the adversarial network;

The detection network training unit is used for training the detection network model, and the feature map of the second image sample used for training uses the trained feature occlusion confrontation network model to add an occlusion mask, and the detection network model is used based on deep learning. image object detection.

Optionally, also include:

The adversarial network retraining unit is used to continue the training of the feature occlusion adversarial network model by using the feature map added with the occlusion mask as an image sample.

Optionally, in the occlusion sample acquisition unit, occlusion of the first image sample includes:

determining the candidate frame with the maximum positioning accuracy in the first image sample;

Map a sliding window of preset size to an image sample, and use background pixels to fill the image area where the sliding window is located;

The detection network model is used to detect the filled image samples, and the sliding window at the position where the candidate frame has the largest detection network loss is used as the occlusion position of the first image sample to obtain the first occlusion image sample.

Optionally, in the training of the detection network model, a non-maximum value suppression algorithm is used to determine the prediction frame, wherein the prediction frame evaluation index in the non-maximum value suppression algorithm is determined by the positioning accuracy and the position accuracy of different candidate frames. Location accuracy is determined.

Optionally, the calculation formula of the candidate frame evaluation index CL is:

CL=γ×sore _class +(1-γ)×score _location ; where γ is the hyperparameter, sore _class is the classification accuracy, and score _location is the positioning accuracy;

The formula of the non-maximum suppression algorithm is:

Among them, CL is the positioning accuracy of different candidate boxes.

Optionally, the method further includes: a preset predicted IoU network model, where the predicted IoU network model is used to obtain the positioning accuracy of different candidate frames.

Optionally, it also includes a prediction IoU network model training unit for generating a candidate frame set of the third image sample; obtaining the positioning accuracy of each candidate frame in the candidate frame set; removing the positioning accuracy in the candidate frame set The candidate frame smaller than the preset threshold is used to determine the training set of the third occlusion image sample; the training set is used to train the prediction IoU network model.

Optionally, the detection network model includes a global pooling module and a pooling classification module, and the pooling classification module is used to classify the globally pooled feature map according to the corresponding position of the pooling mask to obtain a classification. vector.

A computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the instructions are executed on a terminal device, the terminal device is made to execute any one of the above-mentioned images A model building method for object detection.

A computer program product, when the computer program product runs on a terminal device, enables the terminal device to execute any one of the above-mentioned model building methods for image target detection.

The method and device for establishing an image target detection model provided by the embodiments of the present invention use occlusion image samples to train a feature occlusion adversarial network model, and obtain an occlusion mask of an image sample through the feature occlusion adversarial network model. During the training of the model, an occlusion mask is added to the feature map of the image samples used for training, and the occlusion mask is obtained by using the trained feature occlusion confrontation network model. In this way, since the feature occlusion adversarial network model is obtained by training occlusion image samples, a feature occlusion adversarial network model that generates a better mask can be trained, and then the feature occlusion adversarial network is used to obtain an occlusion mask for training the detection network model. Therefore, the detection network model can be trained by using the image samples with better occlusion masks, so that the detection network model can be fully trained for the occlusion situation, and the accuracy of the detection network model for the detection of occluded objects can be improved.

Description of drawings

In order to illustrate the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative effort.

1 shows a flowchart of a method for establishing an image target detection model according to an embodiment of the present invention;

2 shows a schematic diagram of a process for obtaining occlusion samples according to a method for establishing an image target detection model according to an embodiment of the present invention;

3 shows a schematic diagram of a mask according to an embodiment of the present invention;

4 is a schematic structural diagram of an apparatus for establishing an image target detection model according to an embodiment of the present invention.

Detailed ways

In order to make the above objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Many specific details are set forth in the following description to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and those skilled in the art can do so without departing from the connotation of the present invention. Similar promotion, therefore, the present invention is not limited by the specific embodiments disclosed below.

As described in the Background Art, the detection algorithm based on the deep convolutional neural network has become the mainstream method of image target detection, and in the application of image target detection, the object to be detected in the image is often occluded, and the current The detection model seldom considers the occlusion situation and cannot achieve accurate detection of occluded objects. Based on this, this application proposes a method for establishing an image target detection model. During model training, image samples with better occlusion masks are used to train the detection network model, so that the detection network model can be fully trained for the occlusion situation and improve the The accuracy of the detection network model for the detection of occluded objects.

In order to better understand the technical solutions and technical effects of the present application, specific embodiments will be described in detail below with reference to flowcharts.

Referring to FIG. 1 , in step S01, a first image sample is occluded to obtain a first occluded image sample.

The first image sample is an original image sample, and an occluded first image sample is obtained by adding occlusion on the original image sample, that is, the first occlusion image sample.

In a specific application, an appropriate method may be used to perform occlusion. In this embodiment, specifically, the occlusion of the first image sample specifically includes the following steps.

S011, determine a candidate frame with the maximum positioning accuracy in the first image sample.

S012 , map a sliding window of a preset size to an image sample, and use background pixels to fill the image area where the sliding window is located.

S013, the detection network model is used to detect the filled image samples, and the sliding window at the position where the candidate frame has the largest detection network loss is used as the occlusion position of the first image sample, so as to obtain the first occlusion image sample.

For the first image sample, first determine the candidate frame with the maximum positioning accuracy on the first image sample, and the candidate frame with the maximum positioning accuracy has the maximum IoU (Intersection over Union, intersection ratio) with the real mark on the image sample. ) candidate box. The feature map of the image sample can be obtained, different candidate frames are set on the feature map, and the IoU of the different candidate frames and the real mark is calculated, and the candidate frame with the largest IoU value is used as the candidate frame with the maximum positioning accuracy.

Then, set a sliding window with a preset size. The size of the sliding window can be set according to specific needs. For example, it can be a rectangular box of (w/3, d/3) size, and w and d are the length and width of the image sample respectively. , the size of the sliding window will remain unchanged, and the position on the image sample will change. For the sliding window at different positions, map it to the image sample, and use the background pixels to fill the image area where the sliding window is located. The background pixels can be filled randomly so that, at different positions of the image sample, an occlusion is formed by a sliding window.

After that, the detection of the filled image samples is carried out by using the detection network model, and the detection network model is a model for image target detection based on deep learning, that is, a detection network model based on a deep convolutional neural network, which can be, for example, based on YOLO ( You Only Look Once) and SSD (Single Shot MultiBox Detector) and other algorithms, the detection network model can be a trained model. By selecting the sliding window that maximizes the loss of the detection network as the occlusion position of the image sample, the occlusion image sample with the best occlusion position can be obtained.

When the occlusion image sample is obtained by the method of this embodiment, the candidate frame with the maximum positioning accuracy in the image sample is determined, and the position selected by the candidate frame is closest to the real mark, which can better reflect the detection target, and the maximum detection network is used. The sliding window at the position of the loss is used as the sample position of the image, so that the occlusion image sample with the best occlusion position is obtained at the optimal detection target position, and the quality of the occlusion sample is improved.

In order to make it easier to understand the method of obtaining occlusion image samples, the following describes the process of obtaining occlusion samples with an example of an image sample. Referring to Figure 2, first, image samples of multiple candidate frames can be selected in the frame, as shown in Figure 2. As shown in (A), then, the feature map of the image sample is obtained, as shown in Figure (B), and then, the candidate frame with the largest IoU with the real mark is determined, as shown in Figure (C), with the candidate frame The frame is the detection target, and the sliding window is used for occlusion filling, and the sliding window at the position where the candidate frame has the largest detection network loss is used as the occlusion position of the image sample, as shown in Figure (D), in this way, the occlusion position is obtained. occlusion image samples.

In step S02, the first occlusion image sample is used to train a feature occlusion adversarial network model, and the feature occlusion adversarial network model is used to obtain an occlusion mask of the image sample based on the adversarial network.

The feature occlusion adversarial network model is a model for obtaining an occlusion mask of an image sample, which is a model based on an adversarial network and a deep learning model established on the basis of an adversarial network. In this embodiment, the feature occlusion adversarial network is implemented by adding a fully connected layer for learning an occlusion mask after obtaining the feature map of the image sample. The occlusion mask is a mask corresponding to the feature map and is used to reflect Mask to occlude samples.

For ease of understanding, as shown in FIG. 3 , it is a mask M corresponding to one image sample, which is an array mask value corresponding to different positions of the image sample. In addition, it should be noted that, for the convenience of description, in this application, the mask corresponding to the occlusion image sample with occlusion is referred to as an occlusion mask.

Using occlusion image samples to train the feature occlusion adversarial network model, especially high-quality occlusion image samples for training, can obtain a feature occlusion adversarial network model that generates better masks. The loss function of cross entropy can be used, and the specific expression is:

为训练图像样本中第p个图像样本的掩码M中(i,j)位置的值，

为预测值，n表示训练图像样本个数，d表示训练图像样本大小。in,

is the value of the (i,j) position in the mask M of the p-th image sample in the training image sample,

is the predicted value, n represents the number of training image samples, and d represents the training image sample size.

In step S03, the training of the detection network model is performed, and the feature map of the second image sample used for training uses the trained feature occlusion confrontation network model to add an occlusion mask, and the detection network model is used for deep learning-based image targets detection.

As mentioned above, the detection network model is a model based on deep learning and used for image target detection. In this step, the detection network model is continued to be trained, and the training is performed by using a second image sample, and the second image sample can be is a sample different from the above-mentioned first image sample. For the second image sample, after generating the corresponding feature map, add an occlusion mask to the feature map. The occlusion mask is to use the above trained feature to occlude the adversarial network. The model obtained, after adding the occlusion mask, occlusion is performed at the corresponding position of the feature map, that is, the feature map of the second image sample with occlusion is obtained. Using the feature map with occlusion to train the detection network model can enable the detection network model to fully train the occlusion situation and improve the detection accuracy of the detection network model for occluded objects.

In addition, when the detection network model is trained, cross-training of the feature occlusion adversarial network model can be performed. On the one hand, the feature occlusion adversarial network model is used to add occlusion masks to the image samples during the training of the detection network model. The occlusion samples are used to train the detection network model; at the same time, the feature map of the second image sample with the occlusion mask added generated during the training of the detection network model can be used as the image sample, and the training of the feature occlusion confrontation network model can be continued. That is, using the occlusion samples generated during the training of the detection network model, continue to train the feature occlusion adversarial network model, and retraining helps to train a feature occlusion adversarial network model that can generate better masks.

In the training process of the detection network model, the prediction frame needs to be determined from the candidate frame, and the prediction frame evaluation standard plays a decisive role in the accuracy of the prediction frame position. In a more preferred embodiment of the present application, during the training of the detection network model, a non-maximum suppression algorithm (NMS, Non-Maximum Suppression) is used to determine the prediction frame, wherein the non-maximum suppression algorithm in the The prediction frame evaluation index is determined by the positioning accuracy and position accuracy of different candidate frames.

Specifically, the non-maximum value suppression algorithm uses the NMS evaluation index to remove some redundant candidate frames, and determines the prediction frame through correlation calculation in the remaining candidate frames. In this embodiment, the prediction frame evaluation index in the non-maximum value suppression algorithm is determined by the positioning accuracy and position accuracy of different candidate frames. In this way, the classification confidence and positioning accuracy can be comprehensively considered, which is more reasonable. , which is beneficial to improve the prediction accuracy. In a specific application, the calculation formula of the candidate frame evaluation index CL may be:

CL=γ×sore _class +(1-γ)×score _location ; (2)

Among them, γ is a super parameter, sore _class is the classification accuracy, and score _location is the positioning accuracy. The positioning accuracy score _location can be the IoU value, and the classification accuracy sore _class can be obtained by detecting the network model.

In addition, when the candidate frame evaluation index CL is used to determine the prediction frame through the non-maximum value suppression algorithm, different threshold settings will produce different detection results. Reasonable setting of the threshold value can avoid false detection and missed detection. The non-maximum suppression algorithm can be implemented by the following formula with three-stage threshold setting, the specific formula is as follows:

Among them, CL is the positioning accuracy of different candidate frames, and CL can be obtained by the above formula (2).

Through the three-stage threshold setting, for a candidate frame with a large difference, it is basically a frame with low target accuracy. If the threshold is lowered, the candidate frame is easier to be suppressed; and for a candidate frame with a low difference, it is the target. For a frame with high accuracy, if the threshold is increased, the suppression can be reduced, thereby avoiding false detection and missed detection, and improving the detection accuracy.

Further, for the positioning accuracy related to the candidate frame mentioned above, such as when the first image sample is occluded, when the candidate frame with the maximum positioning accuracy in the first image sample is determined, and the training of the detection network model is performed. , the location accuracy used in the prediction frame evaluation index in the non-maximum suppression algorithm can be obtained by using the prediction IoU network model. The predicted IoU network model is used to obtain the positioning accuracy of different candidate boxes. The predicted IoU network model is a deep learning-based model and can include a pooling layer and a fully connected layer.

In a specific application, the training method for predicting the IoU network model may include:

generating a set of candidate boxes for the third image sample;

Obtain the positioning accuracy of each candidate frame in the candidate frame set;

removing candidate frames whose positioning accuracy is less than a preset threshold in the candidate frame set to determine the training set of the third occlusion image sample;

Use the training set to train a predictive IoU network model.

其中，C_i表示第i个候选框，

为第i个候选框与真实标记之间的IoU值，利用该训练集合C进行预测IoU网络模型的训练，在该训练过程中，可以采用平滑损失函数L1，具体表达式为：The third image sample may be the same or different training sample as the first image sample or the second image sample, and frame selection may be performed on the third image sample at random to generate a set of candidate frames for the third image sample, and then, Obtain the positioning accuracy of each candidate frame in the candidate frame set. The positioning accuracy is the IoU value between each candidate frame and the real mark. After deleting the candidate frame with less positioning accuracy, the positioning accuracy is usually less than 0.5. The candidate frame, the training set C of the third occlusion image sample is obtained:

Among them, C _i represents the ith candidate frame,

is the IoU value between the i-th candidate frame and the real mark, and the training set C is used to train the prediction IoU network model. In the training process, the smooth loss function L1 can be used, and the specific expression is:

为第i个候选框与真实标记之间的IoU值。Among them, n represents the number of candidate frames, iou _i is the IoU value between the predicted value of the i-th candidate frame and the true label value,

is the IoU value between the ith candidate box and the ground truth.

In addition, in a specific application process, the candidate frame usually also contains a background image, and the background image will also interfere with object detection, resulting in inaccuracy of detection. In a more preferred embodiment of the present application, the detection network model includes a global pooling module and a pooling classification module, and the pooling classification module is used to classify the globally pooled feature map according to the corresponding position of the pooling mask, to get a categorical vector.

The global pooling module performs global average pooling on the input samples. In the embodiment of this application, the input samples are feature maps with occlusion masks added. The pooling classification module may include a fully connected layer, a pooling mask layer, and a vector In the classification module, the fully connected layer is used to learn the weight of the global average pooling, so as to distinguish the information of different positions, and then obtain the pooling mask of each feature map through the pooling mask layer, and combine the feature map with the pooling mask. Multiply the corresponding positions in the , and then obtain the classification vector, and classify the classification vector through the vector classification module. For example, the vector classification module can be a softmax tool to obtain the classification vector. The classification vector takes into account the background information around the object at the same time, which will have a certain impact on the classification of the object, thereby further improving the detection accuracy.

The model building method for image target detection according to the embodiment of the present application has been described in detail above. In addition, the present application also provides a model building device for image target detection implementing the above method. Referring to FIG. 4 , it includes:

an occlusion sample obtaining unit 400, configured to occlude the first image sample to obtain the first occlusion image sample;

An adversarial network training unit 410, configured to use the first occlusion image sample to train a feature occlusion adversarial network model, where the feature occlusion adversarial network model is used to obtain an occlusion mask of the image sample based on the adversarial network;

The detection network training unit 420 is used for training the detection network model, and the feature map of the second image sample used for training utilizes the trained feature to occlude the confrontation network model with an occlusion mask added, and the detection network model is used for depth-based Learned Image Object Detection.

Further, it also includes:

The adversarial network retraining unit is used to continue the training of the feature occlusion adversarial network model by using the feature map added with the occlusion mask as an image sample.

Further, in the occlusion sample obtaining unit 400, the occlusion of the first image sample includes:

determining the candidate frame with the maximum positioning accuracy in the first image sample;

Map a sliding window of preset size to an image sample, and use background pixels to fill the image area where the sliding window is located;

The detection network model is used to detect the filled image samples, and the sliding window at the position where the candidate frame has the largest detection network loss is used as the occlusion position of the first image sample to obtain the first occlusion image sample.

Further, in the training of the detection network model, a non-maximum value suppression algorithm is used to determine the prediction frame, wherein the prediction frame evaluation index in the non-maximum value suppression algorithm is determined by the positioning accuracy and position of different candidate frames. Accuracy is ok.

Further, the calculation formula of the candidate frame evaluation index CL is:

CL=γ×sore _class +(1-γ)×score _location ; where γ is the hyperparameter, sore _class is the classification accuracy, and score _location is the positioning accuracy;

The formula of the non-maximum suppression algorithm is:

Among them, CL is the positioning accuracy of different candidate boxes.

Further, it also includes: a preset prediction IoU network model, where the prediction IoU network model is used to obtain the positioning accuracy of different candidate frames.

Further, it also includes a prediction IoU network model training unit, which is used to generate a candidate frame set of the third image sample; obtain the positioning accuracy of each candidate frame in the candidate frame set; remove the positioning accuracy in the candidate frame set less than The candidate frame of the preset threshold is used to determine the training set of the third occlusion image sample; the training set is used to train the prediction IoU network model.

Further, the detection network model includes a global pooling module and a pooling classification module, and the pooling classification module is used to classify the global pooled feature map according to the corresponding position of the pooling mask to obtain a classification vector. .

In addition, an embodiment of the present application further provides a computer-readable storage medium, where an instruction is stored in the computer-readable storage medium, and when the instruction is executed on a terminal device, the terminal device is made to execute the above image A model building method for object detection.

Embodiments of the present application further provide a computer program product, which, when running on a terminal device, enables the terminal device to execute the above-mentioned model building method for image target detection.

It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments may be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

It should also be noted that in this document, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply those entities or operations There is no such actual relationship or order between them. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The steps of a method or algorithm described in conjunction with the embodiments disclosed herein may be directly implemented in hardware, a software module executed by a processor, or a combination of the two. A software module can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other in the technical field. in any other known form of storage medium.

The above descriptions are only preferred embodiments of the present invention. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art, without departing from the scope of the technical solution of the present invention, can make many possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above, or modify them into equivalents of equivalent changes. Example. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solutions of the present invention still fall within the protection scope of the technical solutions of the present invention.

Claims (8)

1. a model establishment method of image target detection, is characterized in that, comprises: Determine the candidate frame with the maximum positioning accuracy in the first image sample; map the sliding window of preset size to the image sample, and use the background pixels to fill the image area where the sliding window is located; use the detection network model to fill the image sample , taking the sliding window at the position where the candidate frame has the largest detection network loss as the occlusion position of the first image sample to obtain the first occlusion image sample; Use the first occlusion image sample to train a feature occlusion confrontation network model, where the feature occlusion confrontation network model is used to obtain an occlusion mask of the image sample based on the confrontation network; A detection network model is trained, and an occlusion mask is added to the feature map of the second image sample used for training using the trained feature occlusion confrontation network model, and the detection network model is used for image target detection based on deep learning. 2 . The method according to claim 1 , further comprising: using the feature map added with the occlusion mask as an image sample to continue the training of the feature occlusion adversarial network model. 3 . 3. The method according to claim 1, wherein, in the training of the detection network model, a prediction frame is determined by a non-maximum suppression algorithm, wherein the candidate frame in the non-maximum suppression algorithm The evaluation index is determined by the positioning accuracy and position accuracy of different candidate frames. 4. The method according to claim 3, wherein the calculation formula of the candidate frame evaluation index CL is: CL=γ×sore _class +(1-γ)×score _location ; where γ is the hyperparameter, sore _class is the classification accuracy, and score _location is the positioning accuracy; The formula with three-stage threshold setting adopted by the non-maximum suppression algorithm is:

5. The method according to any one of claims 1, 3-4, wherein a predicted IoU network model is preset, and the predicted IoU network model is used to obtain the positioning accuracy of different candidate frames. 6. The method according to claim 5, wherein the training method of the predicted IoU network model comprises: generating a set of candidate boxes for the third image sample; Obtain the positioning accuracy of each candidate frame in the candidate frame set; removing candidate frames whose positioning accuracy is less than a preset threshold in the candidate frame set to determine the training set of the third occlusion image sample; Use the training set to train a predictive IoU network model. 7. A model establishment device for image target detection, characterized in that, comprising: an occlusion sample acquisition unit, used for determining a candidate frame with the maximum positioning accuracy in the first image sample; mapping a sliding window of a preset size to the image sample, and using background pixels to fill the image area where the sliding window is located; using a detection network The model detects the filled image samples, and uses the sliding window at the position where the candidate frame has the largest detection network loss as the occlusion position of the first image sample to obtain the first occluded image sample; an adversarial network training unit, configured to use the first occlusion image sample to train a feature occlusion adversarial network model, where the feature occlusion adversarial network model is used to obtain an occlusion mask of the image sample based on the adversarial network; The detection network training unit is used for training the detection network model, and the feature map of the second image sample used for training uses the trained feature occlusion confrontation network model to add an occlusion mask, and the detection network model is used based on deep learning. image object detection. 8. A computer-readable storage medium, wherein an instruction is stored in the computer-readable storage medium, and when the instruction is executed on a terminal device, the terminal device is made to execute any one of claims 1-6. The model building method for image target detection described in item.

Images