TW201816661A - Automatic electric scooter identification an d part-based outer defect detection method and system thereof - Google Patents

Abstract

With global warming getting worst, Smart City develops, in which the green transportation plays an important role. However, the public transportation sharing vehicles such as electric scooter and electric bicycle could be damaged by frequent usage, and artificial defect detection system would usually have different inspection standards, and is labor-intensive. In this work, we propose an automatic outer defect detection and identification module for electric scooters at an indoor environment. At first, system provides the hint message of the optimal viewing angle to remind the user capturing the image that is useful for defect detection. The defect localized by feature-based method is decided by comparing the difference between the images captured before rental and after return. The license plate under the optimal viewing angle is also captured for scooter identification. The experimental results from extensive datasets show the robustness of our proposed framework and thus reveal its feasibility for real world environment.

Description

Method and system for automatic motor vehicle identification and vehicle body flaw detection

The invention relates to a method and device for detecting an automatic electric motor vehicle and a vehicle body flaw detection, and proposes an automatic motor vehicle external helium detecting module combined with a vehicle identification module, which can distinguish the motor body part and detect the flaw To clarify the responsibility and to claim the vehicle.

According to the development of science and technology, smart cities manage urban infrastructure through innovative technology and efficiently apply and manage energy resources through analysis of urban infrastructure; Smart cities can enhance the development of cities, and the limited energy can be applied more effectively. In smart cities, intelligent transportation (ITS) has always been a very important piece. In response to global warming, green transportation is promoted. , such as mass transit vehicles, electric car rental systems or electric bicycle rental systems, such as YouBike, are very popular.

After pressing, in the public rental system, the vehicle body will cause multiple damages due to repeated use, which will result in the vehicle not being able to ride, resulting in an increase in the maintenance cost of the vehicle body. Fixed-standard detection systems are often not available on manual inspection systems and can be misjudged due to human error.

The main object of the present invention is to provide an automatic motor vehicle identification and vehicle body detection method, which can capture the current instant shooting angle of view through the action camera to optimize the user's shooting efficiency. It can also ensure the image taken by the user, the object to be tested will not be obscured, and the possible position can be found through the method of feature point positioning to avoid processing the entire image, so as to improve efficiency. It also provides a license plate to calculate the tilt angle by character projection to correct the tilted license plate, and to perform automatic license plate recognition through the word source classifier learned by the support vector machine.

In order to achieve the above object, the method for detecting and detecting vehicle body defects of the present invention comprises: a) capturing an image to be tested, the image to be tested having an original body part image and a license plate image, and setting a Corresponding to the first region of interest of the original body component image and a second region of interest corresponding to the license plate image; b) using a feature extraction algorithm to convert the original body component image of the image to be tested into a feature value, The feature value is utilized by a nearest neighbor search algorithm, and the nearest neighbor search algorithm is established between the original body part image of the image to be tested and the same board image, and a unique set of two-dimensional codes can be generated, and The image similarity unit is formed by determining the degree of similarity between the image of the original body part of the image to be tested and the image of the image through the difference between the two-dimensional codes; c) the image of the original body part of the image to be tested has a plurality of images Positioning points, and grasping the color features of each of the positioning points to form a body part positioning unit; d). the image of the image to be tested has a complex near the license plate image a headlight and converts it into a binary image of the car, and finds the paired features of the lamp in the binarized lamp image to form a pair of lamp units; e). After the combination of the unit, the body component positioning unit and the pair of lamp units, an optimized viewing module is calculated, so that the first region of interest captures the original body component image and the second region of interest The license plate image; f) provides a vehicle body inspection module, which uses a feature 瑕疵 positioning algorithm to search for the 感兴趣 feature of the original body component image captured by the first region of interest, and cooperates with a non-original The image of the body component, and the use of a detection algorithm is different from the image of the original body component and the image of the non-original body component, and finds a feature of the non-original body component image that is not present in the image of the original body component; And g) providing a vehicle identity recognition module that uses a corrected license plate image algorithm, a segmentation character algorithm, and a recognition character algorithm to identify the license plate of the second region of interest The identity on the image.

According to the foregoing feature, the nearest neighbor search algorithm is a KD-ferns algorithm; the feature extraction algorithm is a HOG algorithm step comprising: inputting the original body component image; converting the original body component image into a grayscale image; The gradient direction of the gray-scale image; dividing the gray-scale image into a plurality of cells; and counting the gradient direction histogram of each of the cells; combining the adjacent cells into a block, so that the feature value formed in the block is a HOG feature value.

According to the foregoing feature, the step of capturing the color features of each of the positioning points comprises: converting the original body part image from the RGB color space to the YCbCr color space; and converting the original body part image converted into the YCbCr color space to a predetermined size; The pixel points on the original body part image retracted to the predetermined size are scanned by a mask of a predetermined size, and a part of the pixel points are filtered to a maximum of a plurality of areas; the maximum value of the area is a set distance range, and The maximum value of the area is centered, and is compared with the maximum value of other areas within the distance range. If the maximum value of the area is greater than the maximum value of the other area, the maximum value of the area is set to the maximum value of the range, and vice versa. If the maximum value of the area is smaller than the maximum value of the other area, the maximum value of the area is filtered out, and a plurality of range maximum values are selected; the maximum value of each range is matched with the relative position set by each of the positioning points, and the maximum value of each range is available. Find out where each anchor point is located.

According to the foregoing feature, the feature 瑕疵 positioning algorithm is a SURF algorithm; the 瑕疵 detection algorithm is a projection conversion algorithm.

According to the foregoing feature, the step of correcting the license plate image algorithm comprises: inputting a license plate image having a tilt angle; converting the license plate image into a binarized license plate image; and filtering out a plurality of independent regions set on the binarized license plate image a character within the character; calculating a tilt angle of each of the independent regions, correcting the license plate image of the tilt angle, and calculating a tilt angle of each of the corrected independent regions; if the tilt angle is maintained between -1 and +1, Outputting the corrected license plate image to form a corrected license plate image; otherwise, if the tilt angle is not between -1 and +1, the license plate image regenerates the tilt angle to be corrected; re-correcting the newly generated tilt of the license plate image angle.

According to the foregoing feature, the segmentation character algorithm step comprises: converting the corrected license plate image into a binarized corrected license plate image, and the character on the binarized corrected license plate image is filled with a mask of a predetermined size to form a fill video; determining the symbols to fill the image - can be independently cut; if the symbol - is independently cut, remove the symbol - the position, and the fill video image lines into two portions; conversely, if the symbol - If it cannot be cut independently, the vertical histogram of the position of the midpoint of the image is filled, and the threshold is given, the pixel below the threshold is removed, and the image is divided into two according to the midpoint position. Part; counting the vertical histogram of the filled image, and judging whether there is the existence of the character 1; if there is the presence of the character 1, the appropriate blank space is given before and after the character 1 for cutting; otherwise, if there is no If the character 1 is present, the two partial images are halved and separated, and the different characters are separated. The calculation of each character feature, and the recognition character algorithm is an SVM classifier algorithm step including: Each character recognition is performed using a previously learned SVM classifier.

According to the predecessor feature, the license plate image is a yaw of 46°~65°, a roll of 23°~41° and a pitch of 21°~37° with a viewing angle of the left angle of view. (pitch) to shoot.

In order to achieve the above object, the system for detecting and detecting the vehicle body of the present invention includes: a shooting module that captures an image to be tested, the image to be tested having an original body part image and a license plate image. And setting a first region of interest corresponding to the image of the original body component and a second region of interest corresponding to the license plate image; an optimized viewing module is composed of an image similarity unit, a body component positioning unit, and Calculated after the combination of the lamp unit, the first region of interest captures the original body component image and the second region of interest to capture the license plate image, and the image similarity unit uses a feature extraction The algorithm converts the original body part image of the image to be tested into a feature value, and the feature value is utilized by a nearest neighbor search algorithm, and the nearest neighbor search algorithm is used to establish an image of the original body part of the image to be tested. Between the same panel image, a unique set of two-dimensional codes can be generated, and the original body part of the image to be tested is judged by the difference between the two-dimensional codes. The similarity between the image and the image of the template, and the body component positioning unit has a plurality of positioning points on the image of the original body component of the image to be tested, and captures the color features of each of the positioning points, and then pairs The lamp unit has a plurality of lights in the vicinity of the license plate image of the image to be tested, and is converted into a binarized car light image, and the paired features of the lights are found in the binarized car light image; a vehicle body inspection module, which uses a feature 瑕疵 positioning algorithm to search for the 感兴趣 feature of the original body part image captured by the first region of interest, and cooperates with a non-original body part image, and uses one The 瑕疵 detection algorithm compares the image of the original body part image with the image of the non-original body part, and finds a 瑕疵 feature that the non-original body part image does not present on the original body part image; and a vehicle identity recognition module, The method uses a corrected license plate image algorithm, a segmentation character algorithm and a recognition character algorithm to identify the body of the license plate image captured by the second region of interest. .

By means of the above-mentioned technical means, the method proposed in the invention can be applied to the automatic detection of the external cymbal of the electric motor vehicle, and the automatic license plate recognition is carried out for the license plate with a large inclination angle. Before the 瑕疵 detection, the system will calculate the real-time image information under the shooting to ensure that the shooting angle is at the optimal viewing angle. The optimized viewing angle module has three parts: the image similarity unit and the body part positioning unit (to locate Stickers as positioning targets) and pairs of headlight units. In the automated flaw detection, the feature localization method is used to locate the flaws, and the image differences before and after use are compared to perform the interpretation. In the automated license plate recognition, the projection-based character tilt angle is corrected, and the tilted license plate is turned positive, and the trained character is used to identify the classifier, and the cut characters are recognized.

Firstly, the steps of the preferred embodiment of the method for the identification of the automatic motor vehicle and the detection of the vehicle body of the present invention are as follows. Please refer to the flow chart of FIG. 1 and cooperate with FIG. 2, FIG. 3A-3B, FIG. 4A~4G, FIG. 5A respectively. 5C, Fig. 6, Figs. 7A to 7D, Figs. 8A to 8G, and Figs. 9A to 9B.

Step S1: Shooting an image to be tested (A), the image to be tested (A) has an original body part image (C1) and a license plate image (D), and sets a corresponding image of the original body part (C1) A region of interest (B1) and a second region of interest (B2) corresponding to the license plate image (D) are shown in FIG.

Step S2: Converting the original body part image (C1) of the image to be tested into a feature value by using a feature extraction algorithm, and the feature value is utilized by a Nearest Neighbor Search (NNS) algorithm, and The nearest neighbor search algorithm is established between the original body part image (C1) of the image to be tested (A) and the same board image (X), and a unique set of two-dimensional codes can be generated and transmitted through each of the two-dimensional codes. The difference is the degree of similarity between the original body part image (C1) of the image to be tested (A) and the template image (X), forming an image similarity unit 121, as shown in FIG. 3A, which utilizes a database ( Y) storing a plurality of template images (X), and the nearest neighbor search algorithm is a KD-ferns algorithm for finding each of the template images (X) in the database, and the original body part image (C1) The most similar is the search similarity of the image similarity unit 121, and the characteristic of the KD-ferns algorithm is that under the same layer classification, all nodes have the same branch segmentation value and segmentation dimension, and the classification is selected in each layer. , which can also make the difference in the whole data, The value of the maximum entropy that can be generated after the segmentation is selected from the dimension, and the KD-ferns algorithm refers to the "Fast multiple-part based object detection using KD-Ferns" document, which is a prior art, not the patent of the present invention. Do not repeat them. The feature extraction algorithm is a HOG algorithm comprising: step S21: inputting the original body part image (C1); step S22: converting the original body part image (C1) into a grayscale image; and step S23: calculating the gray level image a gradient direction; step S24: dividing the grayscale image into a plurality of cells; step S25: counting a gradient direction histogram of each of the cells; and step S26: combining the neighboring cells into a block, so that the feature value formed in the block is HOG The feature value, as shown in FIG. 3B, is such that the statistical value of each region in the block is formed into a description element of the image, and the HOG algorithm refers to the document "Histograms of oriented gradients for human detection". It is a prior art, and is not a patent of the present invention.

Step S3: the original body part image (C1) of the image to be tested (A) has a plurality of positioning points (E), and grabs the color features of each of the positioning points (E) to form a body part positioning unit 122; In this embodiment, the step of capturing the color features of each of the positioning points (E) includes: converting the original body part image (C1) from the RGB color space to the YCbCr color space, as shown in FIG. 4A; converting to the YCbCr color. The original body part image (C1) of the space is contracted to a predetermined size as shown in Fig. 4B; the pixel points on the original body part image (C1) retracted to the predetermined size are made with a mask (F) of a predetermined size. After scanning, some pixels are filtered out to be the maximum value (G) of a plurality of regions, as shown in FIG. 4C to FIG. 4D; the maximum value (G) of the region is set to a distance range, and is centered on the maximum value (G) of the region. And within the distance range and outward with other regions maximum value, if the region maximum value (G) is greater than the other region maximum value, the region maximum value (G) is set to the range maximum value (H), On the other hand, if the maximum value (G) of the region is smaller than the maximum value of the other region, the filter is filtered out. The maximum value (G) of the domain is selected, and the maximum value (H) of the plurality of ranges is selected, as shown in FIG. 4E; the maximum value (H) of each of the ranges is matched with the relative position set by each of the positioning points (E), and The position of each of the positioning points (E) is found in the range maximum (H), as shown in FIG. 4F, and finally, as shown in FIG. 4G, is the positioning aspect of the body component positioning unit 122.

Step S4: a plurality of lights (I) are arranged in the vicinity of the license plate image (D) of the image to be tested (A), and converted into a binarized car light image (J), and in the binarized car light image ( J) Finding the paired features of each of the lamps (I) to form a pair of lamp units 123. In this embodiment, Cb and Cr images are generated by Cb and Cr channels in the YCbCr color space (D1, D2) ), as shown in FIG. 5A, the binary image (J) is also generated by giving a threshold, and the position of the lamp (I) is also found, as shown in FIG. 5B, and then the binary value is used. The size of the area of the Unicom area in the image of the vehicle lamp (J) is filtered. Finally, as shown in FIG. 5C, the position of the headlight (I) is captured by the corresponding spatial relationship of the headlights (I). The search for the pair of headlight units 123.

Step S5: After the image similarity unit 121, the body component positioning unit 122, and the pair of headlight units 123 are combined, an optimized view module 12 is calculated to capture the first region of interest (B1). The original body part image (C1) and the second region of interest (B2) capture the license plate image (D), and if the condition of the optimized viewing angle is met, the first region of interest (B1), the second sense The area of interest (B2) is converted to yellow, as shown in Figure 6.

Step S6: providing a vehicle body inspection module 13 for searching for the 瑕疵 feature (N) of the original body part image (C1) captured by the first region of interest (B1) using a feature 瑕疵 positioning algorithm In this embodiment, the feature 瑕疵 positioning algorithm is a SURF algorithm, and the original body component image (C1) is converted by the integral image, the Hessian matrix, and the Hal wavelet, and the SURF feature point (K) is detected. As shown in FIG. 7A, and the SURF algorithm is referred to Speeded-Up Robust Features (SURF), the document is a prior art, not a patent of the present invention, and is not described here, and is accompanied by a non-original body part image. (C2), which uses a detection algorithm to compare the difference between the original body part image (C1) and the non-original body part image (C2). In this embodiment, the detection algorithm is a projection conversion ( Homography) algorithm, and corresponding to the original body part image (C1) and the non-original body part image (C2) relative position, as shown in FIG. 7B, further, find the original body part image (C1) and the non-original Body part image (C2) Different SURF feature points (K) correspond to positions, and the angle of view changes caused by the images taken at different time points are fixed to the same angle of view, so that the non-original body part image (C2) is converted into the original body part image (C1) Converting images (M) of the same viewing angle, and comparing the surrounding areas of different SURF feature points (K) with a predetermined small yellow frame (L), and finding the non-original body part image (C2) in the original body The 瑕疵 feature (N) not shown by the component image (C1), as shown in Fig. 7C, and the 瑕疵 feature (N) of the non-original body component image (C2), is the image of the range to be compared using the HOG feature. Describe, and determine the 瑕疵 feature (N) has the following two steps: (1). Compare the difference between the two corresponding areas of the desired range, and give the threshold, and also determine the candidate who meets the first stage. However, it is impossible to find a very accurate center of the transformation point to perform the transformation matrix, and a slight error is generated, which easily leads to a large number of misjudgments in the first stage; (2) statistical directionality in the comparison area, and the calculation of the directionality Difference, giving a threshold, from And the first-order projective transformation based algorithm to the prior art, non-patentable subject matter of the present invention, capacity is not repeated in the candidates given, also determining the fault signature (N), with FIG. 7D,.

Step S7: providing a vehicle identity recognition module 14 that uses a corrected license plate image algorithm, a segmentation character algorithm, and a recognition character algorithm to identify the license plate image captured by the second region of interest (B2) ( In the embodiment, as shown in FIG. 8A, the step of correcting the license plate image algorithm includes: step S711: inputting a license plate image having a tilt angle (D); and step S712: converting the license plate image (D) a binary image of the license plate image (O); step S713: filtering out the characters in the plurality of independent regions (P) set on the binarized license plate image (O), thus, at the optimal viewing angle At the shooting angle of view, while grabbing the position of the headlight (I) and extending downward from the position of the headlight (I), the approximate position of the license plate can be found, and the license plate image (D) is also obtained. The license plate image (D) is converted into a grayscale image, and the binarized license plate image (O) is generated by giving a bright threshold, and then the independent region (P) is connected in eight directions, and the independent region is calculated and connected ( P) the area size and aspect ratio, also the location of the character, as shown in Figure 8B; step S714: calculation The separate regions (P) of the inclination angle (R & lt), correcting the inclination angle of the license plate (D), and calculates each of the separate regions (P) after the correction of the inclination angle, thus, the first word is calculated The rotation angle between the element and the last character ( θ _rotation ) causes the gray-scale license plate image to form a rotated image (Q) via the rotation angle ( θ _rotation ), as shown in Figure 8C, and gives the threshold of rotation to form one or two. The rotating license plate image (R) is valued and cut by the first character to form a cut binarized license plate image (S), as shown in FIG. 8D, and then, the cut binocular license plate image is searched for ( S) the angle of inclination of each character ( r ), and one character is divided into three parts, as shown in Fig. 8E, the upper half and the lower half are respectively projected on the horizontal axis, for the license plate image without tilt (D), while the upper half of the projection range (up ₀ ) and the lower half of the projection range (down ₀ ) are at the same position, and set the upper half projection range (up ₁ , up ₂₎ with the lower half of the projection range (down _1, down _2), if the character is inclined, the upper half and a lower half Projection midpoint (up _0, down ₀₎ of the inclined projection line, and find a separate area (P) of the cutting plate binarized image (S) in the eight directions through the Unicom zone method, and calculates the individual regions (P The inclination angle ( r ) is shown in Fig. 8F; steps S715, S716: if the inclination angle ( r ) is maintained between -1 and +1, the mode in the inclination angle ( r ) is selected as the cutting The rotation angle ( θ _rotation ) of the binarized license plate image (S) outputs the corrected license plate image (D) to form a corrected license plate image (T), as shown in Fig. 8G; otherwise, steps S715, S717: If the tilt angle ( r ) is not between -1 and +1, the license plate image (D) regenerates the tilt angle to be corrected; and step S718: re-corrects the tilt angle newly generated by the license plate image (D).

In this embodiment, as shown in FIG. 9A, the segmentation character algorithm step includes: step S721: converting the corrected license plate image (T) into a binarized corrected license plate image (W), and the binarized corrected license plate image The character on (W) is filled with a mask (U) of a predetermined size to form a filled image (V), as shown in FIG. 9B; and step S722: determining the symbol on the filled image (V) - No cutting is independent; step S723: If the symbol - is independently cut, remove the symbol - the position, and the filled image (V) is divided into two parts based images; otherwise, step S724: If the symbol - can not be cut independently , the vertical histogram of the front and rear positions of the midpoint of the image (V) is counted, and the threshold is given, the pixel below the threshold is removed, and the filled image (V) is divided according to the midpoint position. Two parts; step S725: counting the vertical histogram of the filled image (V), and judging whether there is the existence of the character 1; step S726: if there is the presence of the character 1, then giving before and after the character 1 Properly blanking space for cutting; otherwise, step S727: if there is no such character 1, the two partial images are Trisected for cutting, independent of different characters; Step S728: calculated for each feature of the character, the image of the next character features determined. After the corrected and cut character image has no fixed shape, the cut character is cut into small areas, and the pixel average value of each small area is calculated, and the average value of each small area is concatenated and used as the character. Image features. The step of identifying the character algorithm as the SVM classifier algorithm comprises: Step S729: performing the character recognition using a well-learned SVM classifier, the SVM classifier algorithm being a prior art, not a patent of the invention The subject matter is not described. In addition, the license plate image (D) is a yaw with a tilt angle tolerance of 46° to 65°, a roll of 23° to 41°, and a pitch of 21° to 37° with a left-angle viewing angle. (pitch) to shoot.

The above-mentioned automatic motor vehicle identification and vehicle body detection method is applied to the system 10 for automatic motor vehicle identification and vehicle body detection. As shown in FIG. 10, it includes: a shooting module 11 that captures an image to be tested ( A), the image to be tested (A) has an original body part image (C1) and a license plate image (D), and sets a first region of interest (B1) corresponding to the original body part image (C1) and a Corresponding to the second region of interest (B2) of the license plate image (D); an optimized viewing module 12 is combined by an image similarity unit 121, a body component positioning unit 122, and a pair of headlight units 123 Calculated so that the first region of interest (B1) captures the original body component image (C1) and the second region of interest (B2) to capture the license plate image (D); The module 13 uses a feature 瑕疵 positioning algorithm to search for the 感兴趣 feature (N) of the original body part image (C1) captured by the first region of interest, and cooperates with a non-original body part image (C2) And it uses a 瑕疵 detection algorithm to compare the original body part image (C 1) distinguishing the non-original body part image (C2) from the non-original body part image (C2) in the original body part image (C1) without the 瑕疵 feature (N); and a vehicle identity recognition mode Group 14, which uses a corrected license plate image algorithm, a segmentation character algorithm, and a recognition character algorithm to identify the identity of the license plate image (D) captured by the second region of interest (B2).

Therefore, the automatic motor vehicle identification and vehicle body detection system 10 is implemented to damage the liability of the rental motor vehicle. When the user takes the image before use, if not used (before use), the definition is: use The person has just completed the rental procedure and has not used it, that is, the photo is stored, and the system 10 for automatic motor vehicle identification and vehicle body detection will calculate the optimized viewing angle immediately when the user takes the image, and wait for the user to take the license plate after the shooting is completed. Identification, after the user confirms, archive according to the license plate identification result, complete the rental procedure, in order to compare with the re-taken image when the user returns; the user needs to perform the second image capture if returned, such as After (after use) is defined as: the user wants to return the leased item, at which time the system 10 of the automatic motor vehicle identification and vehicle body detection system performs flaw detection through an image taken when not in use, the automatic motor The vehicle identification will be the same as the system 10 for detecting the vehicle body. The current view will be optimized and the license will be recognized. After confirmation wearer, according to the recognition result to the license plate read into the database corresponding to the image before capturing of the lease (i.e. the first portion of the image archive), and then the subsequent automatic flaw detection, returned to complete the procedures.

Based on such a composition, the main contributions of the present invention are as follows: (1) Instantaneous optimal viewing angle interpretation; (2) High-performance feature positioning automatic detection; (3) Tilt-based license plate correction based on character projection and automated license plate recognition system . In order to maintain the basic riding function, the automated flaw detection system is very important. The automated flaw detection system can not only significantly reduce the maintenance cost of the vehicle body, but also reduce the errors of human interpretation.

In summary, the structure disclosed by the present invention is unprecedented, and can indeed achieve the improvement of efficacy, and has industrial availability, fully conforms to the patent requirements of the invention, and invites the bureau to grant a patent to encourage innovation. There is no sense of morality.

The drawings and the descriptions of the present invention are merely preferred embodiments of the present invention, and those skilled in the art, which are subject to the spirit of the present invention, should be included in the scope of the patent application.

S1~S7‧‧‧ steps

S21~S26‧‧‧Steps

S711~S718‧‧‧Steps

S721~S729‧‧‧Steps

A‧‧‧ image to be tested

B1‧‧‧First area of interest

B2‧‧‧Second area of interest

C1‧‧‧ original body parts image

C2‧‧‧ original body parts image

D‧‧‧ License Plate Image

D1‧‧‧Cb image

D2‧‧‧Cr image

E‧‧‧ anchor point

F, U‧‧‧ mask

G‧‧‧ area maximum

H‧‧‧Maximum range

I‧‧‧ headlights

J‧‧‧ Binarized car light image

K‧‧‧SURF feature points

L‧‧‧ yellow small box

N‧‧‧瑕疵 characteristics

O‧‧‧ Binarized license plate image

P‧‧‧Independent area

Q‧‧‧Rotating image

R‧‧‧ Binarized rotating license plate image

S‧‧‧ cutting binarized license plate image

T‧‧‧corrected license plate image

V‧‧‧ Filling the image

W‧‧‧ Binarization Correction License Plate Image

X‧‧‧ sample image

Y‧‧‧Database

r ‧‧‧ tilt angle

_Rotation rotation angle of _rotation

Up ₀ ‧‧‧The middle half of the projection range

Down ₀ ‧‧‧The middle half of the projection range

Up ₁ , up ₂ ‧‧‧ upper half projection range

Down ₁ , down ₂ ‧‧‧lower half projection range

10‧‧‧Automatic motor vehicle identification and vehicle body detection system

11‧‧‧ Shooting module

12‧‧‧Optimized viewing module

121‧‧‧Image Similarity Unit

122‧‧‧ Body parts positioning unit

123‧‧‧paired lamp unit

13‧‧‧Car body inspection module

14‧‧‧Vehicle identity recognition module

Figure 1 is a flow chart of the present invention. Figure 2 is a schematic illustration of the capture of the present invention. Figure 3A is a schematic diagram of the KD-fern algorithm search of the present invention. Figure 3B is a schematic illustration of the HOG descriptive element of the present invention. 4A is a schematic diagram of an image conversion color space of an original body component of the present invention. Figure 4B is a schematic illustration of the original body member image of the present invention retracted to a predetermined size. Figure 4C is a schematic illustration of the definition of the region maximum for the present invention. Figure 4D is a schematic illustration of the region maximum search results of the present invention. Figure 4E is a schematic illustration of the results of the maximum range search for the present invention. 4F is a schematic diagram of the search result of the anchor point of the present invention. Fig. 4G is a schematic view showing the image detection result of the original body member of the present invention. FIG. 5A is a schematic diagram of a license plate image of the present invention in a Cr channel and a Cr channel. FIG. 5B is a schematic diagram of the conversion of the license plate image of the present invention into a binarized car light image. Figure 5C is a schematic illustration of the search for paired headlights of the present invention. Fig. 6 is a schematic view showing the image of the original body part and the image of the license plate of the present invention. Figure 7A is a schematic illustration of the SURF feature results of the non-original body part image of the present invention. FIG. 7B is a schematic diagram of the image of the non-original body part and the original body part of the present invention. Figure 7C is a schematic illustration of the comparative step of the present invention. Figure 7D is a schematic illustration of the comparative results of the present invention. Figure 8A is a flow chart of the method of correcting license plate images of the present invention. Figure 8B is a schematic diagram of the license plate image and character position of the present invention. Figure 8C is a schematic illustration of the angle of rotation and the rotated image of the present invention. FIG. 8D is a schematic diagram of the rotating binarized license plate image and the cut binocular license plate image of the present invention. Figure 8E is a schematic illustration of the character projection of the present invention. Figure 8F is a schematic illustration of the angle of inclination exhibited by the individual regions of the present invention. Fig. 8G is a schematic view showing the correction of the license plate image of the present invention. Figure 9A is a flow chart of the license plate recognition of the present invention. Figure 9B is a schematic diagram of the image filling of the present invention. Figure 10 is a view showing the state of use of the present invention.

Claims (10)

The invention relates to an automatic motor vehicle identification and a vehicle body flaw detection method, which comprises: a) capturing an image to be tested, the image to be tested having an original body part image and a license plate image, and setting a corresponding image of the original body part. a first region of interest and a second region of interest corresponding to the license plate image; b) using a feature extraction algorithm to convert the original body component image of the image to be tested into a feature value, the feature value being searched by a nearest neighbor The algorithm is utilized, and the nearest neighbor search algorithm is established between the original body part image of the image to be tested and the same board image, and a unique set of two-dimensional codes can be generated and transmitted through each of the two-dimensional codes. The gap is used to determine the degree of similarity between the original body part image of the image to be tested and the image of the template to form an image similarity unit; c). The original body part image of the image to be tested has a plurality of positioning points and is captured The color features of the positioning points form a body component positioning unit; d). The vehicle image of the image to be tested has a plurality of lights near the license plate image, and is converted a binary image of the lamp, and finding a pair of features of the lamp in the binarized lamp image to form a pair of lamp units; e) passing the image similarity unit and the body component positioning unit And combining the pair of lamp units to calculate an optimized viewing module, so that the first region of interest captures the original body component image and the second region of interest to capture the license plate image; f). Providing a vehicle body inspection module, which uses a feature 瑕疵 positioning algorithm to search for the 感兴趣 feature of the original body part image captured by the first region of interest, and cooperates with a non-original body component image, and uses The detection algorithm compares the image of the original body part image with the image of the non-original body part, and finds a feature of the non-original body part image that is not present in the original body part image; and g) provides a vehicle The identity recognition module uses a corrected license plate image algorithm, a segmentation character algorithm, and a recognition character algorithm to identify the identity of the second region of interest captured on the license plate image. The method for automatic motor vehicle identification and vehicle body flaw detection according to claim 1, wherein the nearest neighbor search algorithm is a KD-ferns algorithm; the feature extraction algorithm is a HOG algorithm step comprising: inputting Converting the original body component image into a grayscale image; calculating a gradient direction of the grayscale image; dividing the grayscale image into a plurality of cells; and counting a gradient direction histogram of each of the cells; combining the adjacent cells into a region The block is such that the feature value formed in the block is a HOG feature value. The method for automatically identifying an electric motor vehicle and detecting a vehicle body according to claim 2, wherein the step of capturing the color features of each of the positioning points comprises: converting the original body part image from the RGB color space to the YCbCr color space. Converting the image of the original body part converted into the YCbCr color space to a predetermined size; the pixel points on the original body part image retracted to the predetermined size are scanned with a mask of a predetermined size, and some pixels are filtered out The point is the maximum value of the plurality of regions; the maximum value of the region is the distance range, and is centered on the maximum value of the region, and is compared with the maximum value of other regions within the distance range, if the maximum value of the region is greater than the other If the maximum value of the area is set, the maximum value of the area is set to the maximum value of the range. Otherwise, if the maximum value of the area is smaller than the maximum value of the other area, the maximum value of the area is filtered out, and the maximum number of the range is selected; The value is matched with the relative position set by each positioning point, and the position of each positioning point can be found from the maximum value of each range. The method for automatic motor vehicle identification and vehicle body flaw detection according to claim 3, wherein the feature 瑕疵 positioning algorithm is a SURF algorithm; the 瑕疵 detection algorithm is a projection conversion algorithm. The method for the automatic motor vehicle identification and the vehicle body detection according to any one of claims 1 to 4, wherein the correcting license plate image algorithm step comprises: inputting a license plate image having an inclined angle; and converting the license plate The image is converted into a binarized license plate image; the characters in the plurality of independent regions set on the binarized license plate image are filtered out; the tilt angle of each of the independent regions is calculated, and the license plate image of the tilt angle is corrected, and the correction is calculated The angle of inclination of each of the independent regions; if the tilt angle is maintained between -1 and +1, the corrected license plate image is output to form a corrected license plate image; otherwise, if the tilt angle is not -1 to +1 Between the two, the license plate image regenerates the tilt angle to be corrected; re-corrects the tilt angle newly generated by the license plate image. The method for automatic motor vehicle identification and vehicle body detection according to claim 5, wherein the segmentation character algorithm step comprises: converting the corrected license plate image into a binarized corrected license plate image, the binary value correction of the character based on the license plate of a predetermined size for mask padding, to fill a form image; determining the symbols to fill the image - can be independently cut; if the symbol - is independently cut, remove the symbol - Position, and divide the image into two partial images; conversely, if the symbol cannot be cut independently, the vertical histogram of the position of the midpoint of the filled image is counted, and the threshold is given, and the valve is removed below The pixel of the value, and according to the midpoint position, divides the padded image system into two parts; counts the vertical histogram of the filled image, and judges whether there is the existence of the character 1; if there is the presence of the character 1, Appropriate blank space is given before and after the character 1 for cutting; otherwise, if there is no such character 1, the two partial images are equally divided into three equal parts, and different characters are independently separated; Each of the character features, and the recognition character algorithm is an SVM classifier algorithm step comprising: performing the character recognition using a previously learned SVM classifier. The method for automatic motor vehicle identification and vehicle body flaw detection according to claim 6, wherein the license plate image is a yaw with a slope tolerance of 46°~65° with a left angle of view. , 23 ° ~ 41 ° roll (roll) and 21 ° ~ 37 ° pitch (pitch) for shooting. A system for automatic motor vehicle identification and vehicle body detection includes: a shooting module that captures an image to be tested, the image to be tested having an original body part image and a license plate image, and setting a corresponding original body a first region of interest of the component image and a second region of interest corresponding to the license plate image; an optimized viewing module is composed of an image similarity unit, a body component positioning unit, and a pair of lamp units Calculated, wherein the first region of interest captures the original body component image and the second region of interest to capture the license plate image, and the image similarity unit converts the to-be-tested using a feature extraction algorithm The image of the original body part of the image is a feature value, and the feature value is utilized by a nearest neighbor search algorithm, and the nearest neighbor search algorithm is established between the image of the original body part of the image to be tested and the same image. Generating a unique set of two-dimensional codes, and determining the original body part image of the image to be tested and the sample image through the difference between the two-dimensional codes The similarity between the two, the body component positioning unit has a plurality of positioning points on the image of the original body component of the image to be tested, and captures the color features of each of the positioning points, and the pair of lamp units are A plurality of lights are arranged in the vicinity of the license plate image of the image to be tested, and converted into a binary image of the lamp, and the paired features of the lights are found in the binarized lamp image; a group that uses a feature 瑕疵 positioning algorithm to search for the 感兴趣 feature of the original body part image captured by the first region of interest, and cooperates with a non-original body part image, and uses a 瑕疵 detection algorithm to compare Determining a difference between the original body part image and the non-original body part image, and finding a non-original body part image that is not present in the original body part image; and a vehicle identity recognition module using a corrected license plate image The algorithm, the segmentation character algorithm and the recognition character algorithm identify the identity of the license plate image captured by the second region of interest. The system for automatic motor vehicle identification and vehicle body detection according to claim 8, wherein the nearest neighbor search algorithm is a KD-ferns algorithm; and the feature extraction algorithm is a HOG algorithm step comprising: Inputting the original body component image; converting the original body component image into a grayscale image; calculating a gradient direction of the grayscale image; dividing the grayscale image into a plurality of cells; and counting a gradient direction histogram of each of the cells; a block, wherein the feature value formed in the block is a HOG feature value, and the step of capturing the color feature of each of the anchor points comprises: converting the original body component image from the RGB color space to the YCbCr color space; The image of the original body part of the YCbCr color space is reduced to a predetermined size; the pixels on the image of the original body part retracted to the predetermined size are scanned by a mask of a predetermined size, and some pixels are filtered out into a plurality of areas. Maximum value; the maximum value of the area is the set distance range, centered on the maximum value of the area, and outward and other areas within the distance range The large value is compared. If the maximum value of the area is greater than the maximum value of the other area, the maximum value of the area is set to the maximum value of the range. Otherwise, if the maximum value of the area is smaller than the maximum value of the other area, the maximum value of the area is filtered out. And selecting a plurality of range maximum values; each of the range maximum values is matched with the relative position set by each of the positioning points, and the position of each positioning point can be found from the maximum value of each range, and then the feature 瑕疵 positioning algorithm For the SURF algorithm; the 瑕疵 detection algorithm is a projection conversion algorithm. The system for automatically identifying an electric motor vehicle and detecting a vehicle body according to claim 9, wherein the step of correcting the license plate image algorithm comprises: inputting a license plate image having a tilt angle; and converting the license plate image into a binary value The license plate image is filtered; the characters in the plurality of independent regions set on the binarized license plate image are filtered; the tilt angle of each of the independent regions is calculated, and the license plate image of the tilt angle is corrected, and each of the corrected independent images is calculated The tilt angle of the area; if the tilt angle is maintained between -1 and +1, the corrected license plate image is output to form a corrected image; otherwise, if the tilt angle is not between -1 and +1, the license plate image Regenerating the tilt angle to be corrected; re-correcting the newly generated tilt angle of the license plate image, and the split character algorithm step includes: converting the corrected license plate image into a binarized corrected license plate image, the binarized corrected license plate image characters based on the size of the image to a predetermined mask padded, padding an image is formed; determining the symbols to fill the image - can be independently cut; if character - independently is cut, remove the symbol - the position, and the fill video image lines into two portions; conversely, if the symbol - can not be independently cut, the count of the vertical position of the longitudinal midpoint of the fill image histogram, And given a threshold, remove the pixel below the threshold, and according to the midpoint position, divide the image into two parts; count the vertical histogram of the filled image, and determine whether there is a character 1 If there is the presence of the character 1, an appropriate blank space is given before and after the character 1 for cutting; otherwise, if there is no such character 1, the two partial images are equally divided into three. And different characters are separately generated; and each character feature is calculated, and the step of the recognized character algorithm being an SVM classifier algorithm includes: using the SVM classifier that has been learned in advance to perform each character recognition, and then the license plate image The yaw, the 23° to 41° roll, and the 21° to 37° pitch are used for the angle of view with a left-angle viewing angle tolerance of 46° to 65°.