TWI592883B - Image recognition system and its adaptive learning method - Google Patents

Description

Image recognition system and its adaptive learning method

The invention relates to an identification system and a learning method thereof, in particular to an image recognition system applied to vehicle image processing and an adaptive learning method thereof.

In recent years, the intelligent Advanced Driver Assistance Systems (ADAS) has developed rapidly, and it is hoped that the use of artificial intelligence will reduce the accident rate of traffic accidents, including lane detection system, reverse assistance system, and vehicle defense in front. Collision systems, etc., are technologies that have been actively developed by domestic and foreign automakers.

In the above system, image recognition technology is an indispensable part of it, and the core of its technology is to use the machine learning algorithm to train the classifier to perform classification and judgment to identify the image. However, due to the limitation of the performance of the in-vehicle embedded system, Therefore, the performance of the classifier on the vehicle has its limit, and how to effectively reduce the false positive rate in the changing road environment under the limited performance of the embedded system of the vehicle is the key target of the current research and development.

Accordingly, a first object of the present invention is to provide an image recognition system that can perform training on its own and efficiently reduce the false positive rate.

Therefore, the image recognition system of the present invention comprises a detecting unit and a learning unit.

The detecting unit captures an image input to obtain an image, and outputs an output image related to the image, and performs arithmetic processing on a set of weak classifier parameters to determine whether the output image meets a preset target, if , an alert signal is output, and the original weak classifier parameter is updated when a new set of weak classifier parameters is received.

The learning unit receives the output image, and performs arithmetic processing on the set of weak classifier parameters and a set of strong classifier parameters to determine whether the output image meets a preset target, and when the two determination results are different, The output image is trained to re-adjust the new weak classifier parameters and update the new weak classifier parameters to the detection unit.

Accordingly, a second object of the present invention is to provide an adaptive learning method for an image recognition system that can perform training on its own and efficiently reduce the false positive rate.

Therefore, the adaptive learning method of the image recognition system of the present invention is applied to the image recognition system as described above, and the method comprises the following steps:

(A) Using the detection unit, an image input is captured to obtain an image and an output image associated with the image.

(B) Using the detecting unit, performing arithmetic processing on a set of weak classifier parameters to determine whether the output image meets a preset target, and if so, generating a warning signal.

(C) Update the original weak classifier parameters when a new set of weak classifier parameters is received.

(D) using the learning unit, performing arithmetic processing on the weak classifier parameter and a set of strong classifier parameters respectively to determine whether the output image meets a preset target, and when the two determination results are different, the output image is The set of new weak classifier parameters is adjusted as a sample to train for step (C) update.

The effect of the invention is that by setting the detecting unit, the image input can be immediately processed, and by setting the learning unit, the output image with the wrong result can be found and trained. In order to reduce the false positive rate, because there is no need for manual auxiliary determination or labeling classification, a large number of training samples can be quickly collected, and the performance can be quickly improved, and the false positive rate can be reduced to increase the driver's driving safety.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same reference numerals.

Referring to FIG. 1 and FIG. 2, a first embodiment of the image recognition system of the present invention comprises a detecting unit 2 and a learning unit 3.

The detecting unit 2 is adapted to be disposed in a vehicle 9 to capture an image input to obtain an image, and output an output image related to the image, and perform operation processing on a set of weak classifier parameters to determine the output image. If the image meets the preset target, if it is met, a warning signal is output, and when a new weak classifier parameter is received, the original weak classifier parameter is updated.

In the embodiment, the image input is an image taken by a camera (not shown) for photographing the direction of the vehicle, so that the detecting unit 2 determines whether the direction of the vehicle is obstructed. The preset target is the obstacle preset by the system, but the detecting unit 2 can also be applied to other vehicle image processing, and is not limited thereto.

The detecting unit 2 includes an image capturing module 21, a first image processing module 22, a first sorting module 23, an output module 24, and a parameter module 25.

The image capturing module 21 captures the image input output by the camera for the camera to output the image.

The first image processing module 22 is electrically connected to the image capturing module 21, receives the image, and performs processing to output a first region image.

The first classification module 23 is electrically connected to the first image processing module 22, receives the first area image, and performs arithmetic processing on the set of weak classifier parameters to determine the first area map corresponding to the output image. If the compliance with the preset target is met, if the vehicle direction has an obstacle, the first classification module 23 outputs the warning signal to alert the driver.

The output module 24 is electrically connected to the image capturing module 21 for receiving the image and outputting the output image.

The parameter module 25 is configured to store the set of weak classifier parameters for use by the first classification module 23 for calculation.

The learning unit 3 is connected to the detecting unit 2, receives the output image, and performs arithmetic processing on the weak classifier parameter and a set of strong classifier parameters to determine whether the output image meets a preset target. When the two determination results are different, the output image is trained to readjust the new weak classifier parameters, and the new weak classifier parameters are updated to the parameter module 25.

In this embodiment, the learning unit 3 can be disposed in the vehicle or separately provided in a server (not shown), and is wired or wirelessly connected to the detecting unit 2 for the parameter module 25 Update the new weak classifier parameters for this group.

The learning unit 3 includes a second image processing module 31, a second classification module 32, a training module 33, and an update module 34.

The second image processing module 31 receives the output image and performs arithmetic processing to output the output image and a second region image.

The second classification module 32 receives the output image and the second region image, and performs arithmetic processing on the set of weak classifier parameters and the set of strong classifier parameters to determine the second corresponding to the output image. Whether the area image meets the preset target, since the second classification module 32 and the first classification module 23 use the same set of weak classifier parameters for calculation, therefore, the operation of the group of weak classifier parameters is used here. The determination result can be regarded as the same as the determination result calculated by the first classification module 23, and since the number of the strong classifier parameters is large, the accuracy is high, and therefore, the calculation is performed by using the strong classifier parameters of the group. The determination result may be regarded as a correct determination result, and when the determination result of using the set of weak classifier parameters and using the set of strong classifier parameters is different, it indicates that the determination result calculated by the first classification module 23 is incorrect. At this time, the output image corresponding to the determination result is output.

The training module 33 receives the output image corresponding to the determination result by the second classification module 32, and trains the second classification module 32 with the output image as a sample to adjust the new weak classification of the group. Parameters.

The update module 34 receives the new weak classifier parameters from the training module 33, and updates the new weak classifier parameters to the second classification module 32 and the parameter module 25, thereby The embodiment can find the output image corresponding to the determination result incorrectly, and retrain itself to adjust the weak classifier parameter, and then automatically update the parameter.

The update module 34 tests and calculates a trust score of one of the new weak classifier parameters by using a preset set of image samples to confirm that the set of new weak classifier parameters is an obstacle preset by the system. Whether the judgment performance of (for example, a person, a car, etc.) is better than the original weak classifier parameter, and when the trust score of the new weak classifier parameter is higher than the trust score of the original weak classifier parameter, the new weak group is updated. The classifier parameters are to the second classification module 32 and the image capturing module 21.

It is to be noted that the first classification module 23 can have a weak classifier 231, and the weak classifier 231 can perform arithmetic processing with the weak classifier parameters, and the second classification module 32 can have As shown in FIG. 3, a hierarchical strong classifier 321 is used, and the strong classifier 321 is matched with the strong classifier parameters for processing, and the M class of the strong classifier 321 forms a weak classifier 322. The arithmetic processing is performed in conjunction with the set of weak classifier parameters, whereby the second classification module 32 can output two determination results respectively.

Through the above description, the advantages of this embodiment can be summarized as follows:

1. By setting the detecting unit 2, the image input output by the camera for the vehicle can be immediately subjected to arithmetic processing, and when the output image is determined to meet the preset target (when there is an obstacle), the warning signal is output. The user is alerted, and by continuously updating the set of weak classifier parameters, the classification accuracy of the first classification module 23 can be continuously improved to reduce the false positive rate.

Furthermore, the learning unit 3 connected to the detecting unit 2 is provided, and the second sorting module 32 performs arithmetic processing on the weak classifier parameters and the set of strong classifier parameters respectively. The result of the two determinations may determine whether the determination result calculated by the first classification module 23 is incorrect, and when the determination result is incorrect, the training module 33 is trained to readjust the weak classifier parameter, the update mode. The group 34 updates the new weak classifier parameters of the group, whereby the embodiment can find the output image with the wrong determination result and perform self-training to reduce the false positive rate, since the operation is automatically in the system. It does not need to manually perform auxiliary judgment or label classification, so it can quickly collect a large number of training samples for training to get better set of weak classifier parameters, so it can quickly improve performance and reduce false positive rate to increase driver's driving safety. .

The learning unit 3 can be provided by arranging the detecting unit 2 on the vehicle embedded system on the vehicle 9 and setting the learning unit 3 on the vehicle 9 or the server to provide wired or wireless connection. The second classification module 32, which is powerful in computation, can be constructed without being limited by the performance of the in-vehicle embedded system, so that the output image can be calculated more quickly and instantaneously to provide the group of the first classification module 23 better. Weak classifier parameters.

2. By updating the set of new weak classifier parameters with a higher trust score at a time, the update module 34 can ensure better performance for each update, and avoid the output image that is misjudged at the time. When it is only for special events, it will result in poor detection and misjudgment.

Referring to FIG. 4 and FIG. 5, a second embodiment of the image recognition system of the present invention is similar to the first embodiment. The difference between the second embodiment and the first embodiment is as follows:

The first classification module 23 further includes a distance calculation unit 232. When the weak classifier 231 determines that the vehicle direction has an obstacle, the distance calculation unit 232 receives the first area image according to the first area map. The obstacle size in the image calculates the distance of the obstacle to output a distance signal indicating the distance of an obstacle.

The output module 24 outputs the output image to the learning unit 3 when receiving a false positive feedback signal.

The second embodiment further includes a feedback unit 4 that receives a body signal, an impact signal, and the distance signal.

In this embodiment, the body signal indicates a vehicle speed value and a brake pedal depth value, but may be at least one of the indications, and is not limited thereto.

The feedback unit 4 outputs the false positive feedback signal when the following four situations occur:

The feedback unit 4 receives the warning signal, and the deceleration change amount of the vehicle speed value indicated by the body signal is not greater than a deceleration predetermined value or the brake pedal depth value is not greater than a brake predetermined value.

In this case, the first classification module 23 determines that there is an obstacle (according to the preset target), but the vehicle 9 does not decelerate rapidly or the driver does not step on the brake. Therefore, the feedback unit 4 determines the first classification module. 23 judgment is wrong.

The feedback unit 4 does not receive the warning signal, and the deceleration change amount of the vehicle speed value indicated by the body signal is greater than the deceleration predetermined value or the brake pedal depth value is greater than the brake predetermined value.

This indicates that the first classification module 23 does not judge that there is an obstacle (not meeting the preset target), but the vehicle 9 decelerates rapidly or the driver repeatedly brakes on the brake, indicating that the first classification module 23 determines that there is an error.

3. The feedback unit 4 receives the impact signal.

This indicates that the first classification module 23 does not output the warning signal or outputs the warning signal too late, causing the driver to react and collide.

4. When the obstacle distance is less than a calculated braking distance.

In this case, the first classification module 23 determines that there is an obstacle (according to the preset target) within the safe braking distance (ie, the calculated braking distance) based on the obstacle distance calculated by the distance calculating unit 232, but driving However, the driver is still driving close, so that the first classification module 23 is judged to be in error.

Where the braking distance is calculated Calculated according to the following formula, Set to 0 for the scheduled end speed. For the initial speed, that is, the speed value of the vehicle, For acceleration, calculated by the brake force (0.4g) specified by the European New Car Assessment Programme (Euro-NCAP), For the theoretical braking distance calculated according to Equation 1, For vehicle displacement during the reaction, Calculated by the driver's reaction time of about 0.8 seconds. (Formula 1) (Formula 2) (Formula 3)

Thus, the second embodiment can achieve the same purpose and efficacy as the first embodiment described above, and can also achieve the following advantages:

1. By setting the feedback unit 4 and causing the output module 24 to output the output image to the learning unit 3 when receiving the false positive feedback signal, the output image output to the learning unit 3 can be greatly reduced. Therefore, the amount of information processing of the second classification module 32 can be greatly reduced, that is, the requirement for the performance of the second classification module 32 can be reduced, and the design of the second classification module 32 can be simplified. Therefore, the second classification module 32 can be implemented using a general computing architecture, or the learning unit 3 can be directly disposed with the detection unit 2 in the vehicle embedded system of the vehicle 9, thereby The design cost is saved, and the cost of separately setting the learning unit 3 on the vehicle can be saved, or the communication unit 3 can be connected to the server, and the connection can be updated without being limited to having a communication signal. The first classification module 23 is provided with instant parameter feedback.

2. By providing the warning signal, the body signal, the distance signal and the impact signal to the feedback unit 4, the feedback unit 4 can be used to determine whether the first classification module 23 is misjudged, and to help control the output module. The output images output by the 24 are effective training samples corresponding to the judgment error, so as to reduce the information processing amount of the second classification module 32 and save the design cost.

Referring to FIG. 4 and FIG. 6, the adaptive learning method of the image recognition system executed by the image recognition system includes the following steps 51-56.

Step 51: Using the detecting unit 2, an image input is captured to obtain an image and an output image related to the image.

The image is processed to generate a first area image, and the detecting unit 2 outputs the output image to the learning unit 3 when receiving a false positive feedback signal.

Step 52: Using the detecting unit 2, performing arithmetic processing on a set of weak classifier parameters to determine whether the output image meets a preset target, and if so, generating an alert signal.

The operation process is performed by using the weak classifier parameter to determine whether the first region image corresponding to the output image meets the preset target, and if yes, the alert signal is generated.

Step 53: Update the original weak classifier parameters when a new set of weak classifier parameters is received.

Step 54: Using the learning unit 3, performing arithmetic processing on the set of weak classifier parameters and a set of strong classifier parameters to determine whether the output image meets a preset target, and when the two determination results are different, the output is The image is trained as a sample to adjust the new set of weak classifier parameters for step 53 update.

The output image is processed to generate a second region image, and the operation is processed by the set of weak classifier parameters and the set of strong classifier parameters to determine the second region corresponding to the output image. If the two determination results are different, the output image is used as a sample to train and adjust the new weak classifier parameters for updating in step 53.

Wherein, the trust score of one of the new weak classifier parameters is calculated, and when the trust score of the new weak classifier parameter is higher than the trust score of the original weak classifier parameter, the new weak classifier parameter is output. Updated for step 53.

Step 55: Determine whether the false positive feedback signal is generated based on at least a body signal and the warning signal.

Wherein, whether the false positive feedback signal is generated is also determined according to a distance signal indicating an obstacle distance.

The body signal indicates at least one of a vehicle speed value and a brake pedal depth value, and the false positive feedback signal is generated in the following four cases:

1. Receiving the warning signal, and the deceleration change amount of the vehicle speed value indicated by the body signal is not greater than a deceleration predetermined value or the brake pedal depth value is not greater than a brake predetermined value.

2. The warning signal is not received, and the deceleration change amount of the vehicle speed value indicated by the body signal is greater than the deceleration predetermined value or the brake pedal depth value is greater than the brake predetermined value.

Third, when receiving an impact signal.

4. When the obstacle distance is less than a calculated braking distance, wherein the calculated braking distance is calculated according to the driving speed value.

Thus, the adaptive learning method can achieve the same purpose and effect as the first embodiment described above.

In summary, by setting the detecting unit 2 and the learning unit 3, the image input can be immediately processed, and the output image with the wrong result can be found for training, which can quickly improve the performance. The false positive rate is reduced to increase the driver's driving safety, so the object of the present invention can be achieved.

However, the above is only the embodiment of the present invention, and the scope of the invention is not limited thereto, and all the equivalent equivalent changes and modifications according to the scope of the patent application and the patent specification of the present invention are still The scope of the invention is covered.

2‧‧‧Detection unit
21‧‧‧Image capture module
22‧‧‧First Image Processing Module
23‧‧‧First classification module
231‧‧‧Weak classifier
232‧‧‧ Distance Calculation Department
24‧‧‧Output module
25‧‧‧Parameter Module
3‧‧‧Learning unit
31‧‧‧Second image processing module
32‧‧‧Second classification module
321‧‧‧strong classifier
322‧‧‧Weak classifier
33‧‧‧ training module
34‧‧‧Update Module
4‧‧‧Feedback unit
51~55‧‧‧Steps
9‧‧‧ Vehicles

Other features and effects of the present invention will be apparent from the following description of the drawings. FIG. 1 is a circuit block diagram of a first embodiment of the image recognition system of the present invention; FIG. 3 is a schematic diagram of a strong classifier and a weak classifier of the first embodiment; FIG. 4 is a circuit block diagram of a second embodiment of the image recognition system of the present invention; FIG. 5 is a schematic diagram of the second embodiment applied to a vehicle; and FIG. 6 is a flowchart of an adaptive learning method of the image recognition system of the present invention.

2‧‧‧Detection unit

21‧‧‧Image capture module

22‧‧‧First Image Processing Module

23‧‧‧First classification module

231‧‧‧Weak classifier

232‧‧‧ Distance Calculation Department

24‧‧‧Output module

3‧‧‧Learning unit

31‧‧‧Second image processing module

32‧‧‧Second classification module

321‧‧‧strong classifier

33‧‧‧ training module

34‧‧‧Update Module

4‧‧‧Feedback unit

25‧‧‧Parameter Module

Claims (13)

An image recognition system includes: a detecting unit that captures an image input to obtain an image, and outputs an output image related to the image, and performs arithmetic processing on a set of weak classifier parameters to determine the output image Whether the preset target is met, if yes, a warning signal is output, and when a new weak classifier parameter is received, the original weak classifier parameter is updated; and a learning unit receives the output image and respectively The set of weak classifier parameters and a set of strong classifier parameters are processed to determine whether the output image meets a preset target. When the two determination results are different, the output image is trained to readjust the new group. The weak classifier parameter and update the new weak classifier parameter to the detection unit. The image recognition system of claim 1, wherein the detecting unit comprises: an image capturing module that captures the image input to output the image, and a first image processing module electrically connected to the image capturing device The module receives the image and performs processing to output a first area image, a first classification module, electrically connects the first image processing module, receives the first area image, and uses the group of weak classifiers The parameter is processed to determine whether the image of the first area corresponding to the output image meets the preset target, and if yes, the warning signal is output, and an output module is electrically connected to the image capturing module for The image is received and the output image is output. The image recognition system of claim 2, wherein the learning unit comprises: a second image processing module, receiving the output image and performing processing to output the output image and a second region image, The second classification module receives the output image and the second area image, and performs arithmetic processing on the set of weak classifier parameters and the set of strong classifier parameters to determine the second area map corresponding to the output image Outputting the output image when the two determination results are different, a training module, receiving the output image by the second classification module, and training the second image with the output image as a sample The classification module adjusts the new weak classifier parameters of the group, and an update module, the training module receives the new weak classifier parameters, and updates the group of new weak classifier parameters to the second Classification module and the detection unit. The image recognition system of claim 3, further comprising a feedback unit, the feedback unit receiving a body signal, and outputting a false positive feedback signal according to at least the body signal, the output module outputting the error feedback signal when receiving the error signal Output image. The image recognition system of claim 4, wherein the body signal indicates at least one of a vehicle speed value and a brake pedal depth value, and the feedback unit outputs the false positive feedback signal when at least one of the following occurs: The feedback unit receives the warning signal, and the deceleration change amount of the vehicle speed value indicated by the body signal is not greater than a deceleration predetermined value or the brake pedal depth value is not greater than a brake predetermined value; the feedback unit does not receive The warning signal, and the deceleration change amount of the vehicle speed value indicated by the vehicle body signal at a unit time is greater than the deceleration predetermined value or the brake pedal depth value is greater than the brake predetermined value. The image recognition system of claim 5, wherein the feedback unit further receives an impact signal and a distance signal indicating an obstacle distance, and outputs the false positive feedback signal when at least one of the following occurs: the feedback unit Receiving the impact signal, when the obstacle distance is less than a calculated braking distance, wherein the calculated braking distance is calculated according to the vehicle speed value. The image recognition system of claim 3, wherein the update module calculates a trust score of one of the new weak classifier parameters, and the trust score of the new weak classifier parameter is higher than the original weak classifier When the parameter has a trust score, the new weak classifier parameter of the group is updated to the second classification module and the detecting unit. An adaptive learning method for an image recognition system is applied to the image recognition system of claim 1, the method comprising the following steps: (A) using the detection unit to capture an image input to obtain an image and a correlation The output image of the image; (B) using the detecting unit, performing arithmetic processing on a set of weak classifier parameters to determine whether the output image meets a preset target, and if so, generating a warning signal; Updating the original weak classifier parameter when receiving a new set of weak classifier parameters; and (D) using the learning unit to perform arithmetic processing on the weak classifier parameter and a set of strong classifier parameters respectively Whether the output image meets the preset target, when the two determination results are different, the output image is used as a sample to train and adjust the new weak classifier parameter for the step (C) to update. The adaptive learning method of claim 8, wherein: in step (A), the image is processed to generate a first region image, and in step (B), the weak classifier parameter is used for arithmetic processing. Determining whether the first area image corresponding to the output image meets the preset target, if yes, generating the warning signal, and in step (D), calculating the output image to generate a second area image And performing arithmetic processing on the set of weak classifier parameters and the set of strong classifier parameters to determine whether the image of the second region corresponding to the output image meets the preset target, and when the two determination results are different, The output image is used as a sample to train to adjust the new set of weak classifier parameters for step (C) update. The adaptive learning method of claim 9, wherein in the step (A), the detecting unit outputs the output image to the learning unit when receiving a false positive feedback signal. The adaptive learning method of claim 10, further comprising the following steps: (E) determining, according to at least one body signal and the warning signal, whether the false positive feedback signal is generated, wherein the body signal indicates a vehicle speed value and a The brake pedal depth value is at least one of the following, and the false positive feedback signal is generated in at least one of the following conditions: receiving the warning signal, and the deceleration change amount of the vehicle speed value indicated by the body signal at the unit time is not greater than a deceleration predetermined The value or the brake pedal depth value is not greater than a predetermined brake value, the warning signal is not received, and the deceleration change amount of the vehicle speed value indicated by the body signal at the unit time is greater than the deceleration predetermined value or the brake pedal depth value. Greater than the predetermined value of the brake. The adaptive learning method of claim 11, wherein in step (E), determining whether the false positive feedback signal is generated according to a distance signal indicating an obstacle distance, and generating the false positive feedback signal in at least one of the following conditions; : When receiving an impact signal, when the obstacle distance is less than a calculated braking distance, wherein the calculated braking distance is calculated according to the driving speed value. The adaptive learning method according to claim 9, wherein in step (D), one of the new weak classifier parameters is calculated to have a trust score, and the trust score of the new weak classifier parameter is higher than the original When the trust score of the weak classifier parameter is obtained, the new weak classifier parameter of the group is output for step (C) update.