JP7222209B2 - DEEP LEARNING NETWORK USED FOR EVENT DETECTION, TRAINING DEVICE AND METHOD FOR THE NETWORK - Google Patents

Description

TECHNICAL FIELD The present invention relates to the field of information technology, and more particularly to a deep learning network used for event detection, a training device for the network, and a training method.

Recently, deep learning has been widely applied in the field of computer vision. Deep learning is moving the direction of research in the field of computer vision from image classification to video analysis, such as event detection. Compared to image classification, video analysis faces more complex scenes, and event detection also requires models to learn higher-level logical judgments.

FIG. 1 is a diagram showing detection results of a conventional event detection model. As shown in FIG. 1, on the monitoring video screen shown in FIG. 1, the probability of occurrence of each event detected by the conventional event detection model is 0.03 for normal, 0.46 for accident, and 0.46 for traffic congestion. Jam) 0.41, Park 0.08, and Reverse 0.02.

It should be noted that the above description of the technical background is provided for a clear and complete understanding of the technical solution of the present invention, and is provided for the understanding of those skilled in the art. These technical ideas are merely described as part of the background art of the present invention and are not well known by those skilled in the art.

According to the discovery of the inventor of the present invention, in the case of FIG. Then, the output result cannot be obtained, and if the output condition is the output of the event with the maximum probability, the output result is an accident. That is, the conventional event detection model can only distinguish mutually exclusive events, cannot output multiple events as detection results, and can ensure the accuracy and completeness of detection results. Can not. In addition, since the conventional event detection model is a multi-classifier detection model, its training time is long.

Embodiments of the present invention provide a deep learning network used for event detection, an apparatus for training the network and a training method. The deep learning network has at least two event classifiers that detect different events independently of each other, and each event classifier detects independently and outputs the detection result, so that the event detection result Accuracy and completeness can be ensured. Also, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

In a first aspect of an embodiment of the present invention, a deep learning network used for event detection, comprising: a data layer reading input data; a convolutional layer extracting features from the input data read by the data layer; at least two event classifiers that detect different events independently of each other based on the features extracted by the convolutional layers and output detection results of the different events, respectively. .

According to a second aspect of an embodiment of the present invention, a training apparatus for a deep learning network according to the first aspect of an embodiment of the present invention, comprising first training means for training parameters of said convolutional layers of said deep learning network. and second training means for training the parameters of the at least two event classifiers of the deep learning network while maintaining the parameters of the convolutional layers of the deep learning network.

In a third aspect of an embodiment of the present invention, a method for training a deep learning network according to the first aspect of an embodiment of the present invention, comprising the steps of training parameters of said convolutional layers of said deep learning network; and training the parameters of the at least two event classifiers of the deep learning network while maintaining the parameters of the convolutional layers of the deep learning network.

Advantageously, the deep learning network has at least two event classifiers for detecting different events independently of each other, each event classifier for independently detecting and generating a detection result of By outputting, the accuracy and completeness of event detection results can be ensured. Also, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

Specific embodiments of the invention are disclosed in detail, as set forth in the following description and drawings, to illustrate the manner in which the principles of the invention may be employed. It should be noted that embodiments of the present invention are not limited in scope. Embodiments of the present invention include various changes, modifications and equivalents within the spirit and content of the appended claims.

Features described and/or shown in one embodiment may be used in one or many other embodiments in the same or similar manner and may be combined with features in other embodiments; Features in other embodiments may be substituted.

It should be noted that the term "comprising/including" as used herein means the presence of a feature, element, step or component, and the presence or absence of one or more other features, elements, steps or components. Additions are not excluded.

The drawings included herein are included to provide an understanding of embodiments of the invention, and constitute a part of this specification and are intended to illustrate embodiments of the invention and, together with the written description, The principle of the present invention will now be explained. It should be noted that the drawings described herein are merely for explaining the embodiments of the present invention, and those skilled in the art can easily obtain other drawings based on these drawings.
FIG. 10 is a diagram showing detection results of a conventional event detection model; 1 is a diagram showing a deep learning network used for event detection in Example 1 of the present invention; FIG. It is a figure which shows the detection result of the deep learning network of Example 1 of this invention. It is a figure which shows the event classifier 203 of Example 1 of this invention. Fig. 2 is a diagram showing a training device according to Example 2 of the present invention; It is a figure which shows the electronic device of Example 3 of this invention. FIG. 10 is a block diagram showing the system configuration of an electronic device according to Example 3 of the present invention; It is a figure which shows the training method of Example 4 of this invention.

These and other features of the invention can be understood from the drawings and the description below. The specification and drawings disclose specific embodiments of the invention, which represent some embodiments consistent with the principles of the invention. It should be noted that the invention is not limited to the described embodiments, but that the invention includes all modifications, variations and equivalents that come within the scope of the claims.

<Example 1>
Embodiments of the present invention provide deep learning networks for use in event detection. FIG. 2 is a diagram showing a deep learning network used for event detection in Example 1 of the present invention. As shown in FIG. 2, deep learning network 200 includes data layer 201 , convolutional layer 202 and at least two event classifiers 203 .

Data layer 201 reads input data.

A convolutional layer 202 extracts features from the input data read by the data layer.

At least two event classifiers 203 independently detect different events based on the features extracted by the convolutional layers, and output different event detection results.

According to the above embodiment, the deep learning network has at least two event classifiers that detect different events independently of each other, and each event classifier independently detects and outputs a detection result. By doing so, the accuracy and completeness of event detection results can be ensured. Also, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

In this example, data layer 201 reads input data. For example, the data layer 201 processes surveillance video and obtains input data.

For example, the input data may be at least one frame of surveillance video, which may be captured by a surveillance camera installed above the road.

In this embodiment, convolutional layer 202 extracts features from the input data read by the data layer. The convolutional layer 202 may use conventional structures. For example, the convolutional layer 202 may be a conventional Alexnet network structure.

In this embodiment, the feature may be each feature in the surveillance video image as the input data, such as contour, texture, brightness, and the like.

In this embodiment, at least two event classifiers 203 independently detect different events based on the features extracted by the convolutional layer 202, and respectively output different event detection results.

In this embodiment, each event classifier 203 can detect different events, and each event classifier 203 can only detect one kind of event, ie each event classifier 203 is a dichotomous classifier.

In this embodiment, the number of event classifiers 203 may be set according to actual demand. For example, it may be set according to the number of types of events to be detected.

For example, as shown in FIG. 2, the deep learning network 200 has five events for detecting each event: Normal, Accident, Jam, Park and Reverse. An event classifier 203 may be included.

In this embodiment, detection results output by at least two event classifiers 203 may be displayed. For example, different event detection results output by at least two event classifiers 203 may be collectively displayed on the surveillance video screen.

FIG. 3 is a diagram showing detection results of the deep learning network of Example 1 of the present invention. As shown in FIG. 3, on a monitoring video screen similar to that in FIG. 1, the detection results obtained by the deep learning network 200 for a similar input video are 0.01 for Normal and 0.01 for Accident. 96, Jam 0.89, Park 0.31, Reverse 0.10. Thus, the deep learning network 200 has five event classifiers 203 for detecting normal (Normal), accident (Accident), jam (Jam), parking (Park) and reverse (Reverse), By independently detecting different events by each event classifier 203, the accuracy and completeness of event detection results can be ensured.

In this embodiment, the structure of each event classifier 203 may be the same or different. In this embodiment, the event classifier 203 having the same structure will be described as an example.

FIG. 4 is a diagram showing the event classifier 203 of Example 1 of the present invention. As shown in FIG. 4 , the event classifier 203 includes a first fully connected layer 401 , a second fully connected layer 402 , and a long short-term memory layer provided between the first fully connected layer 401 and the second fully connected layer 402 . (LSTM: Long Short-Term Memory) layer 403 .

In this embodiment, by providing an LSTM layer in the event classifier, it is possible to obtain high detection accuracy by using the characteristic of storing useful information over time and forgetting useless information.

In this embodiment, the event classifier 203 may further include an output layer 404 for outputting the probability of occurrence of events detected by the event classifier 203 .

In this embodiment, the first fully-connected layer 401, the second fully-connected layer 402, the LSTM layer 403, and the output layer 404 may all have conventional structures.

In this embodiment, each event classifier 203 can be trained independently and/or its parameters adjusted independently. In this way, the deep learning network 200 can be flexibly trained and/or tuned, effectively reducing training and/or tuning time.

In this embodiment, event classifiers 203 can be independently added to or deleted from the deep learning network 200 .

For example, when detecting a new event according to the actual situation, an event classifier for detecting the new event may be added to the deep learning network 200 independently. If event detection becomes unnecessary according to the actual situation, the event classifier for detecting the event may be deleted from the deep learning network 200 .

In this way, the deep learning network can have flexible expansion and deletion functions, and the event classifiers in the deep learning network can be increased or decreased according to actual demands.

According to the above embodiment, the deep learning network has at least two event classifiers that detect different events independently of each other, and each event classifier independently detects and outputs a detection result. By doing so, the accuracy and completeness of event detection results can be ensured. Also, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

<Example 2>
An embodiment of the present invention further provides an apparatus for training a deep learning network used for event detection as described in the first embodiment. The structure of the deep learning network is shown in FIG. 2, the deep learning network 200 includes a data layer 201, a convolutional layer 202 and at least two event classifiers 203. FIG.

FIG. 5 is a diagram showing a training device of embodiment 2 of the present invention. As shown in FIG. 5, training device 500 includes first training section 501 and second training section 502 .

A first training unit 501 trains the parameters of the convolutional layer 202 of the deep learning network 200 .

A second training unit 502 trains the parameters of at least two event classifiers 203 of the deep learning network 200 while maintaining the parameters of the convolutional layers 202 of the deep learning network 200 .

Thus, each event classifier needs to detect only one type of event, i.e. each event classifier is a dichotomous classifier, thus requiring less training time.

In this embodiment, the first training unit 501 trains parameters of the convolutional layer 202 of the deep learning network 200 .

For example, public data sets may be used to train the parameters of convolutional layer 202 . Thus, since the public dataset contains more than one million images, the model can provide rich features for training parameters, and the trained model has good universality. .

In the present example, training may be performed in a convolutional neural network structure (Caffe: Convolutional Architecture for Fast Feature Embedding), in which a normal Alexnet network is followed by two full We need to add a coupling layer, one accuracy layer and one loss layer. In the training process, based on the output values of the accuracy layer and the loss layer, it is determined whether the model has converged, and if converged, the training is terminated. After the training is completed, we delete the added two fully connected layers, one accuracy layer and one loss layer to obtain the trained convolutional layer 202 .

In this embodiment, after the training of the convolutional layer 202 is completed, the second training unit 502 performs at least two event classifications of the deep learning network 200 while maintaining the parameters of the convolutional layer 202 of the deep learning network 200. 203 parameters are trained.

For example, the training may be performed using acquired monitoring video data. During training, one accuracy layer and one loss layer need to be added, and in the training process, the learning rate of the convolutional layer 202 is set to 0, i.e. without changing the parameters of the convolutional layer 202 maintain. Further, the learning rate of event classifiers that do not require training among the event classifiers 203 may be set to zero. In the training process, based on the output values of the accuracy layer and the loss layer, it is determined whether the model has converged, and if converged, the training is terminated. After training is completed, we remove the added accuracy and loss layers to obtain a trained event classifier 203 .

In this embodiment, the second training unit 502 uses at least two labels represented by binarized numerical values respectively corresponding to the at least two event classifiers, Parameters may be trained.

For example, the label "1" may indicate that an event has occurred, and the label "0" may indicate that an event has not occurred. ), congestion (Jam), parking (Park), and reverse (Reverse) events. For example, in the surveillance image screen shown in FIG. 1, the label corresponding to each event classifier 203 may be expressed as "01100".

In this embodiment, the second training unit 502 may train the parameters of the at least two event classifiers 203 simultaneously, or may train the parameters of each event classifier 203 of the at least two event classifiers respectively. .

In this way, if the parameters of each event classifier 203 are trained simultaneously, the training time can be further reduced, and if the parameters of each event classifier 203 are trained separately, the training can be flexible according to the actual situation. can be done.

In this embodiment, event classifiers 203 can be independently added to or deleted from the deep learning network 200 . Therefore, the training device 500 may further include a third training section 503 .

When a new event classifier 203 is added to the deep learning network 200, the third training unit 503 maintains the parameters of the convolutional layer 202 and the original at least two event classifiers 203 while adding the new event classifier 203 to the deep learning network 200. It trains the parameters of the new event classifier 203 that have been created by itself. The specific training method may refer to the training method of the original event classifier, and the description thereof is omitted here.

Thus, when there is a new detection request, the convolutional layer 202 and the original at least two event classifiers 203 need not be retrained, but the parameters of the new event classifier 203 can be trained alone. Therefore, the training time can be effectively reduced, and new detection requirements can be quickly met. Also, the training process of the new event classifier does not affect the original event classifier, thus ensuring the detection accuracy of the original event classifier.

In this example, training device 500 may further include adjustment unit 504 .

The adjusting unit 504 independently adjusts parameters of one or more event classifiers of the at least two event classifiers 203 if one or more event classifiers do not meet a predetermined condition.

In this embodiment, the predetermined condition is, for example, that the detection accuracy of the event classifier has reached a predetermined threshold. For example, in the process of performing event detection using the deep learning network, one or more event The detection accuracy of the classifier 203 has decreased and has become lower than the predetermined threshold. In this way, by independently adjusting the parameters of the one or more event classifiers 203 using the adjustment unit 504, the other event classifiers that do not need adjustment are not affected, and various situations can be achieved. adjustments can be made flexibly and quickly.

For example, the tuning process adds features to the original training data that have not been used in these previous trainings, and independently trains the event classifier or classifiers 203 that need to be tuned, Once training is complete, tuning the parameters of the one or more event classifiers 203 is complete.

According to the above embodiments, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

<Example 3>
Embodiments of the present invention further provide an electronic device, and FIG. 6 is a schematic diagram of the electronic device of Embodiment 3 of the present invention. As shown in FIG. 6, the electronic device 600 includes a training device 601, which trains the deep learning network described in the first embodiment. The configuration and function of the training device 601 are the same as those described in Example 2, and the description thereof is omitted here.

FIG. 7 is a block diagram showing the system configuration of an electronic device according to Example 3 of the present invention. As shown in FIG. 7, the electronic device 700 may include a central processing unit (central control unit) 701 and a storage device 702 , the storage device 702 being connected to the central processing unit 701 . The diagrams are merely exemplary and other types of structures may be used to supplement or replace the structures to implement telecommunications or other functions.

As shown in FIG. 7, electronic device 700 may further include input 703 , display 704 and power supply 705 .

In one aspect, the functionality of the training device of Example 2 may be integrated into central processing unit 701 . Here, the central processing unit 701 trains the parameters of the convolutional layers of the deep learning network, and while maintaining the parameters of the convolutional layers of the deep learning network, trains the parameters of at least two event classifiers of the deep learning network. may be configured to train the

For example, training parameters of the at least two event classifiers includes training parameters of the at least two event classifiers simultaneously or training parameters of each event classifier of the at least two event classifiers respectively. step.

For example, when an event classifier is added to the deep learning network, the central processing unit 701 maintains the parameters of the convolutional layer and the at least two event classifiers while maintaining the event classifier added to the deep learning network. It may be configured to train the classifier parameters alone.

For example, central processing unit 701 may independently adjust parameters of one or more event classifiers of the at least two event classifiers if one or more event classifiers do not meet a predetermined condition. may be configured to

For example, training the parameters of the at least two event classifiers includes using at least two binarized numerical labels respectively corresponding to the at least two event classifiers to obtain the at least two training the parameters of the two event classifiers.

In another aspect, the training devices described in Example 2 may each be configured with a central processing unit 701, e.g. may implement the functionality of the training device.

The electronic device 700 in this embodiment does not have to include all the components shown in FIG.

As shown in FIG. 7, central processing unit 701, also referred to as a controller or operational control unit, which may include a microprocessor or other processing and/or logic device, central processing unit 701 receives inputs and processes electronic It controls the operation of each part of the device 700 .

Storage device 702 may be, for example, one or more of a buffer, flash memory, hard disk, removable media, volatile memory, non-volatile memory, or other suitable device. Also, the central processing unit 701 may execute programs stored in the storage device 702 to realize storage or processing of information. Other members are similar to the prior art, so their description is omitted here. Each portion of electronic device 700 may be implemented by specific hardware, firmware, software, or a combination thereof without departing from the scope of the invention.

According to this embodiment, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

<Example 4>
An embodiment of the present invention further provides a training method for the deep learning network used for event detection in embodiment 1, which training method corresponds to the training apparatus in embodiment 2. FIG. 8 is a diagram showing a training method according to embodiment 4 of the present invention. As shown in FIG. 8, the method includes the following steps.

Step 801: Train the parameters of the convolutional layers of the deep learning network.

Step 802: Training the parameters of the at least two event classifiers of the deep learning network while maintaining the parameters of the convolutional layers of the deep learning network.

In this embodiment, the method may further include the following steps.

Step 803: If an event classifier has been added to the deep learning network, while maintaining the parameters of the convolutional layer and the at least two event classifiers, set the parameters of the event classifier added to the deep learning network. train alone.

Step 804: Adjust parameters of the one or more event classifiers independently if one or more of the at least two event classifiers do not meet a predetermined condition.

In this embodiment, the specific implementation method of each of the above steps is the same as that described in Embodiment 2, and the description thereof is omitted here.

According to this embodiment, since each event classifier needs to detect only one type of event, the time required for training the deep learning network is short, and the trained deep learning network has high detection accuracy.

An embodiment of the present invention provides a method for executing a program in a deep learning network training device or an electronic device, causing a computer to perform the deep learning network training described in the above embodiment 4 in the deep learning network training device or electronic device. Further provided is a computer readable program for carrying out the method.

An embodiment of the present invention includes a storage medium storing a computer-readable program for causing a computer to execute the deep learning network training method according to the fourth embodiment in a deep learning network training apparatus or electronic device. Offer more.

The deep learning network training method performed in the deep learning network training device or electronic device described with reference to the embodiments of the present invention may be implemented in hardware, software modules executed by a processor, or a combination of both. may For example, one or more of the functional block diagrams shown in FIG. 5, or one or more combinations of functional block diagrams, may correspond to each software module of the computer program flow, or to each hardware module. You can respond. These software modules may respectively correspond to the steps shown in FIG. These hardware modules may be implemented by hardwareizing these software modules using, for example, a field programmable gate array (FPGA).

A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, mobile hard disk, a CD-ROM, or any other form of storage medium known to those skilled in the art. The storage medium may be coupled to the processor such that the processor reads information from and writes information to the storage medium, and the storage medium may be a component of the processor. The processor and storage medium are located in the ASIC. The software module may be stored in the memory of the mobile terminal or may be stored on a memory card inserted into the mobile terminal. For example, if the device (eg, mobile terminal) uses a relatively large-capacity MEGA-SIM card or a large-capacity flash memory device, the software module is stored in the MEGA-SIM card or large-capacity flash memory device. good too.

One or more of the functional blocks and/or one or more combinations of functional blocks in the functional block diagram depicted in FIG. DSP), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or any suitable combination thereof. may be One or more of the functional blocks and/or one or more combinations of functional blocks in the functional block diagram depicted in FIG. It may be implemented in a combination of processors, one or more microprocessors in combination with DSP communications, or any other configuration.

Although the present invention has been described with reference to specific embodiments, the above description is merely illustrative and does not limit the scope of protection of the present invention. Various variations and modifications may be made to the present invention without departing from the spirit and principle of the present invention, and these variations and modifications are also within the scope of the present invention.

In addition, the following notes are further disclosed with respect to the embodiments including the above-described examples.
(Appendix 1)
A deep learning network used for event detection, comprising:
a data layer that reads input data;
a convolutional layer that extracts features from the input data read by the data layer;
at least two event classifiers that detect different events independently of each other based on the features extracted by the convolutional layers and output different event detection results, respectively.
(Appendix 2)
2. The deep learning network of clause 1, wherein the at least two event classifiers have the same or different structures.
(Appendix 3)
the at least two event classifiers have the same structure;
3. The method of claim 2, wherein the event classifier includes a first fully connected layer, a second fully connected layer, and a long-term memory layer provided between the first fully connected layer and the second fully connected layer. deep learning network.
(Appendix 4)
2. The deep learning network of clause 1, wherein each said event classifier can be independently trained and/or independently parameter tuned.
(Appendix 5)
Clause 1. The deep learning network of Clause 1, wherein the event classifier can be independently added to or deleted from the deep learning network.
(Appendix 6)
A deep learning network training device according to Supplementary Note 1,
a first training means for training parameters of the convolutional layers of the deep learning network;
and second training means for training the parameters of the at least two event classifiers of the deep learning network while maintaining the parameters of the convolutional layers of the deep learning network.
(Appendix 7)
7. Apparatus according to clause 6, wherein said second training means trains parameters of said at least two event classifiers simultaneously or trains parameters of each event classifier of said at least two event classifiers respectively.
(Appendix 8)
If an event classifier is added to the deep learning network, training the parameters of the event classifier added to the deep learning network alone while maintaining the parameters of the convolutional layer and the at least two event classifiers. 7. The apparatus of clause 6, further comprising a third training means for:
(Appendix 9)
adjusting means for independently adjusting parameters of the one or more event classifiers if one or more of the at least two event classifiers does not meet a predetermined condition; 6. Apparatus according to clause 6.
(Appendix 10)
The second training means trains parameters of the at least two event classifiers using at least two labels represented by binarized numerical values respectively corresponding to the at least two event classifiers. 6. Apparatus according to clause 6.
(Appendix 11)
A method of training a deep learning network according to Supplementary Note 1, comprising:
training parameters of the convolutional layers of the deep learning network;
training parameters of the at least two event classifiers of the deep learning network while maintaining parameters of the convolutional layers of the deep learning network.
(Appendix 12)
training the parameters of the at least two event classifiers,
12. The method of clause 11, comprising training parameters of the at least two event classifiers simultaneously, or training parameters of each event classifier of the at least two event classifiers, respectively.
(Appendix 13)
If an event classifier is added to the deep learning network, training the parameters of the event classifier added to the deep learning network alone while maintaining the parameters of the convolutional layer and the at least two event classifiers. 12. The method of clause 11, further comprising the step of:
(Appendix 14)
independently adjusting parameters of the one or more event classifiers if one or more of the at least two event classifiers do not meet a predetermined condition. 11. The method according to 11.
(Appendix 15)
training the parameters of the at least two event classifiers,
training parameters of the at least two event classifiers using at least two binarized numerical labels respectively corresponding to the at least two event classifiers. described method.

Claims (10)

A deep learning network used for event detection, comprising:
a data layer that reads input data;
a convolutional layer that extracts features from the input data read by the data layer;
at least two event classifiers that detect different events independently of each other based on the features extracted by the convolutional layers and output different event detection results, respectively;
A deep learning network , wherein each said event classifier detects only one event . 2. The deep learning network of claim 1, wherein the at least two event classifiers have identical or different structures. the at least two event classifiers have the same structure;
3. The event classifier of claim 2, wherein the event classifier includes a first fully connected layer, a second fully connected layer, and a long-term memory layer provided between the first fully connected layer and the second fully connected layer. deep learning network. 2. The deep learning network of claim 1, wherein each said event classifier can be independently trained and/or independently parameter tuned. 2. The deep learning network of claim 1, wherein the event classifiers can be independently added to or removed from the deep learning network. A deep learning network training device according to claim 1,
a first training means for training parameters of the convolutional layers of the deep learning network;
and second training means for training the parameters of the at least two event classifiers of the deep learning network while maintaining the parameters of the convolutional layers of the deep learning network. 7. Apparatus according to claim 6, wherein said second training means trains parameters of said at least two event classifiers simultaneously or trains parameters of each event classifier of said at least two event classifiers respectively. If an event classifier is added to the deep learning network, training the parameters of the event classifier added to the deep learning network alone while maintaining the parameters of the convolutional layer and the at least two event classifiers. 7. The apparatus of claim 6, further comprising a third training means for: adjusting means for independently adjusting parameters of the one or more event classifiers if one or more of the at least two event classifiers does not meet a predetermined condition; 7. Apparatus according to claim 6. The second training means trains the parameters of the at least two event classifiers using at least two labels represented by binarized numerical values respectively corresponding to the at least two event classifiers. 7. Apparatus according to claim 6.

Images