TWI780409B - Method and system for training object detection model - Google Patents

Abstract

The invention provides a method and a system for training an object detection model. The method includes: obtaining a target object image and a labelled result, wherein the target object image includes a target object, and the labelled result includes a tag corresponding to the target object; obtaining a specific image captured by a fixed camera for a specific scene; obtaining a background image corresponding to a specific image based on the specific image; synthesizing the target object and the background image into a training image; training a target object detection model exclusive for the fixed camera based on the training image and the labelled result corresponding to the target object.

Description

Method and system for training object detection model

The present invention relates to a model training technique, and in particular to a method and system for training an object detection model.

With the rapid development of cameras, network technology, and artificial intelligence, the application of intelligent video surveillance technology has grown significantly. In the application of monitoring system, the detection of specific types of objects (such as people, vehicles, etc.) is a core function. Existing methods for detecting specific types of objects in images, including the use of deep neural network-like techniques, learn the characteristics of target types of objects through training with a large amount of labeled data.

In general, labeled data contains only positive samples (types of objects to be detected), while negative samples are automatically obtained from the background of the training images. However, the diversity of negative samples is limited by the training data. Therefore, for variable images, it is often misjudged that the background is a target type object, which leads to false detection.

In addition, in order to be able to judge extremely diverse negative samples, models with good recognition effects tend to be complex and large, requiring more computing resources in use. The limitation caused by this in actual use is that the camera device cannot afford detection Due to the computational complexity of the model, the image data must be transmitted to the server through the network for detection, resulting in a burden on the network bandwidth and the computational complexity of the server.

Generally speaking, the difficulties in the use of existing technologies partly stem from the requirement for the versatility of the object detection model, that is, the same model can be used in various images and scenes. For the application of video surveillance, each camera is responsible for the scene, and the false detections that are likely to occur in different scenes are different. Detection is quite repetitive.

Therefore, if each camera can have its own detection model, and the detection model has been trained for the scene to avoid false detection of the scene, then the simpler model can be used to achieve the required detection performance, and Improve the feasibility of placing the object detection function on the camera side instead of the server side (that is, the concept of edge computing), and reduce the demand for server computing power and network bandwidth.

In view of this, the present invention provides a method and system for training an object detection model, which can be used to solve the above technical problems.

The present invention provides a method for training an object detection model, comprising: obtaining a target object image and a marking result, wherein the target object image includes at least one target object, and the marking result includes a mark corresponding to each target object; obtaining a first object A fixed camera shoots at least one first specific image of a first specific scene; obtains a first background image corresponding to the at least one first specific image according to the at least one first specific image; combines at least one target object with the first The background image is synthesized into a first training image; a first object detection model dedicated to the first fixed camera is trained based on the first training image and the marking results corresponding to each object.

The invention provides a system for training an object detection model, which includes a data subsystem and a training subsystem. The data subsystem is configured to: obtain an object image, wherein the object image includes at least one object and a labeling result corresponding to each object; obtain at least one image of a first specific scene captured by a first stationary camera. The first specific image; obtaining a first background image corresponding to the at least one first specific image according to the at least one first specific image; synthesizing at least one target object and the first background image into a first training image. The training subsystem trains a first object detection model dedicated to the first fixed camera based on the first training image and the marking results corresponding to each object.

Please refer to FIG. 1 , which is a schematic diagram of a system for training an object detection model according to an embodiment of the present invention. As shown in FIG. 1 , the system 10 includes a detection subsystem 200 , a training subsystem 300 and a data subsystem 400 . In different embodiments, the detection subsystem, the training subsystem 300, and the data subsystem 400 can be realized by independent devices/devices (such as personal circuits, servers, workstations, etc.), or integrated into a single device / device, but the present invention is not limited thereto.

In the embodiment of the present invention, the detection subsystem 200 , the training subsystem 300 and the data subsystem 400 can cooperate to implement the method for training the object detection model proposed by the present invention. In a nutshell, the method proposed by the present invention can generate new training data based on general positive data and negative data specific to a certain fixed camera, and use this training data to train the target specific to the fixed camera The object detection model, the relevant details will be detailed later.

For the convenience of description, the first fixed camera 100 in FIG. 1 will be taken as an example for description below. In the embodiment of the present invention, the first fixed camera 100 is, for example, fixedly installed at a specific location and configured to shoot a first specific scene 199 based on a fixed imaging range.

Please refer to FIG. 2 , which is a flowchart of a method for training an object detection model according to an embodiment of the present invention. The method of this embodiment can be executed by the system 10 in FIG. 1 , and the details of each step in FIG. 2 will be described below with the components shown in FIG. 1 .

Firstly, in step S210, the data subsystem 400 can obtain the image of the target object and the marking result. For ease of understanding, Figure 3 is supplemented below for illustration.

Please refer to FIG. 3 , which is a schematic diagram of a target object image drawn and a marking result according to an embodiment of the present invention. In this embodiment, the target object image 310 may, for example, only include the target objects 310 a - 310 e (such as people) without including other image components (such as background). In addition, the tagged result 320 corresponding to the target object image 310 may include tags 320a-320e corresponding to the target objects 310a-310e.

In the embodiment of the present invention, the target objects 310 a - 310 e in the target object image 310 can be understood as the aforementioned positive data, but the present invention is not limited thereto. In other embodiments, the target object image 310 (and the labeling result 320 ) can be used in addition to training the first target object detection model M1 dedicated to the fixed camera 100 with the negative data corresponding to the fixed camera 100 , and can also be used for training object detection models specific to other stationary cameras. In other words, the target object image 310 (and the labeling result 320 ) can be widely used to train a target object detection model of multiple fixed cameras, and the relevant details will be described later.

After that, in step S220 , the data subsystem 400 can obtain the first specific image captured by the first fixed camera 100 on the first specific scene 199 . Afterwards, in step S230 , the data subsystem 400 can obtain a first background image corresponding to the first specific image according to the first specific image. For ease of understanding, Figure 4 is supplemented below for illustration.

Please refer to FIG. 4 , which is a schematic diagram of a first specific image and its corresponding first background image according to an embodiment of the present invention. In this embodiment, the first specific image 410 is, for example, an image captured by the first fixed camera 100 on the first specific scene 199 . In different embodiments, the first fixed camera 100 may regularly or irregularly send the captured image as the first specific image 410 to the data subsystem 400 to reduce the related bandwidth requirement, but is not limited thereto.

After obtaining a plurality of first specific images 410 corresponding to the same scene (ie, the first specific scene 199 ), the data subsystem 400 can obtain images corresponding to The first background image 420 of the aforementioned one or more first specific images 410 . As shown in FIG. 4, the obtained first background image 420 only includes background objects (such as objects 420a~420c) in the first specific scene 199 and does not include foreground objects and/or target objects (such as people), but may not limited to this. In the embodiment of the present invention, the first background image 420 can be understood as the aforementioned negative data, but the present invention is not limited thereto.

Afterwards, in step S240 , the data subsystem 400 can synthesize the target objects 310 a - 310 e and the first background image 420 into a first training image, and the first training image can be understood as including both positive data and negative data. Next, in step S250 , the training subsystem 300 can train the first target object detection model M1 dedicated to the first fixed camera 100 based on the first training image and the labeling results 320 corresponding to the target objects 310 a - 310 e . For ease of understanding, FIG. 5 is supplemented below for illustration.

Please refer to FIG. 5 , which is a schematic diagram of the first training image and corresponding labeling results shown in FIG. 3 and FIG. 4 . In this embodiment, the first training image 510 is obtained, for example, by inserting the target objects 310 a - 310 e into the first background image 420 by the data subsystem 400 , but it is not limited thereto. In some embodiments, when the data subsystem 400 combines the target objects 310a-310e with the first background image 420, relevant data enhancement processing, such as scaling, rotation, translation, and color adjustment, can be performed on the target objects 310a-310e at the same time. , partial cropping, etc.

In this way, the data subsystem 400 can generate training data specific to the first fixed camera 100 , and the training subsystem 300 can train the first target object specific to the first fixed camera 100 with the help of the corresponding labeling results 320 Detection model M1.

Furthermore, different from the conventional object detection model with higher versatility, the first target object detection model M1 trained by the method of the present invention is specially used to detect objects appearing in the first specific scene 199 The target object (such as a person) can effectively improve the relevant detection performance and reduce the probability of false detection.

In one embodiment, in order to reduce the bandwidth requirement, the detection subsystem 200 can also be configured in the first fixed camera 100 , but it is not limited thereto. After the training subsystem 300 completes the training of the first object detection model M1 , the training subsystem 300 can transmit the first object detection model M1 to the detection subsystem 200 . Afterwards, the detection subsystem 200 can obtain the first image IM1 shot by the first fixed camera 100 on the first specific scene 199, and use the first target object detection model M1 to detect the target object appearing in the first image IM1 (such as a character). In some embodiments, the detection subsystem 200 can generate a corresponding detection result after detecting the target object in the first image IM1, which can include information such as the type, size, position, outline, and confidence value of the target object. , but not limited to this.

In other embodiments, the first image IM1 can be further used as the aforementioned first specific image 410 to further optimize the obtained first background image 420 . Afterwards, the data subsystem 400 can synthesize the optimized first background image 420 with target objects in other target object images to form other training images for the training subsystem 300 to further train the first target object detection model M1, In order to obtain the first object detection model M1 with better detection performance, but the present invention is not limited thereto.

In addition, as mentioned earlier, the system 10 of the present invention can also train a target object detection model specific to other fixed cameras based on the target object image 310 . For example, the data subsystem 400 may be configured to: obtain a second specific image shot by a second fixed camera on a second specific scene; obtain a second background image corresponding to the second specific image according to the second specific image; The target object and the second background image are synthesized into a second training image. Afterwards, the training subsystem 300 can train a second object detection model dedicated to the second fixed camera based on the second training image and the labeling results corresponding to each object. Relevant details of the above technical means can refer to the descriptions in the previous embodiments, and will not be repeated here.

In summary, the system and method proposed by the present invention can establish corresponding detection models for different scenes without manually marking data for different scenes, which is easy to use and can effectively improve related detection models. Test performance, and reduce the probability of false detection.

Although the present invention has been disclosed above with the embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the technical field may make some changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be defined by the scope of the appended patent application.

100: The first fixed camera 199: The first specific scene 200: Detection Subsystem 300: training subsystem 310: target object image 310a~310e: target object 320: mark the result 320a~320e: mark 400: Data Subsystem 410: The first specific image 420: The first background image 420a~420c: object 510: First training image IM1: first image M1: The first object detection model S210~S250: steps

FIG. 1 is a schematic diagram of a system for training an object detection model according to an embodiment of the present invention. FIG. 2 is a flowchart of a method for training an object detection model according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a target object image drawn and a marking result according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a first specific image and its corresponding first background image according to an embodiment of the present invention. FIG. 5 is a schematic diagram of the first training image and corresponding labeling results shown in FIG. 3 and FIG. 4 .

S210~S250: steps

Claims (8)

A method for training an object detection model, comprising: obtaining a target object image and a marking result, wherein the target object image includes at least one target object, and the marking result includes a mark corresponding to each of the target objects; obtaining a first A fixed camera to capture at least one first specific image of a first specific scene based on a fixed imaging range; obtain a first background image corresponding to the at least one first specific image according to the at least one first specific image ; synthesize the at least one target object and the first background image into a first training image; train a first fixed camera dedicated to the first fixed camera based on the first training image and the marking results corresponding to each of the target objects A target object detection model, and based on the target object image training a second target object detection model specific to a second fixed camera, the steps include: obtaining a second specific scene shot by the second fixed camera at least one second specific image; obtain a second background image corresponding to the at least one second specific image according to the at least one second specific image; and synthesize the at least one target object and the second background image into a first Two training images. The method as claimed in claim 1, wherein the target object image only includes the at least one target object. The method as described in claim 1, further comprising: obtaining a first image of the first specific scene shot by the first fixed camera; detecting objects appearing in the first image using the first target object detection model The at least one target object of . The method as described in claim 1, wherein the step of training the second target object detection model specific to the second fixed camera based on the target object image further includes: based on the second training image and corresponding to each of the targets The labeled results of objects train the second object detection model specific to the second stationary camera. A system for training an object detection model, comprising: a data subsystem configured to: obtain an image of a target object and a labeling result, wherein the image of the target object includes at least one target object, and the labeling result includes information corresponding to each A mark of the target object; obtain at least one first specific image taken by a first fixed camera to a first specific scene based on a fixed imaging range; obtain corresponding to the at least one specific image according to the at least one first specific image A first background image of the first specific image; combining the at least one target object and the first background image into a first training training image; a training subsystem, which is based on the first training image and the labeling results corresponding to each of the target objects to train a first target object detection model specific to the first fixed camera, and based on the target object Image training for a second target object detection model dedicated to a second fixed camera, the steps of which include: obtaining at least one second specific image shot by the second fixed camera on a second specific scene; according to the at least one The second specific image obtains a second background image corresponding to the at least one second specific image; and synthesizes the at least one target object and the second background image into a second training image. The system as claimed in claim 5, wherein the target object image only includes the at least one target object. The system as described in claim 5 further includes a detection subsystem configured to: acquire a first image of the first specific scene shot by the first fixed camera; detect the first target object The model detects the at least one target object appearing in the first image. The system of claim 5, wherein the training subsystem is further configured to train the second target object specific to the second stationary camera based on the second training image and the labeling results corresponding to each of the target objects detection model.

Images