CN111881702A

CN111881702A - Multi-code rapid reading method based on YOLO-barQR algorithm

Info

Publication number: CN111881702A
Application number: CN202010693776.1A
Authority: CN
Inventors: 章漠松; 梁步亮
Original assignee: Niu Niu Tu Technology Co ltd
Current assignee: Niu Niu Tu Technology Co ltd
Priority date: 2020-07-17
Filing date: 2020-07-17
Publication date: 2020-11-03

Abstract

The invention discloses a multi-code quick reading method based on a YOLO-barQR algorithm, which comprises the following steps: step S1: acquiring a picture of a graph to be identified; step S2: adopting a YOLO-barQR network model after the data set picture training to identify all codes in the picture; step S3: adopting the YOLO-barQR network model to position the coordinates of each code and evaluating the quality weight of each code; step S4: and sequencing according to the quality weight, sequentially cutting the graph of each code according to the coordinates to obtain pictures, and outputting the pictures to a decoding module for decoding. The method can identify all codes in the picture at one time, accurately position each code, does not depend on the resolution of the picture, does not need to research the characteristics of each code system, and saves a large amount of research and development investment.

Description

Multi-code rapid reading method based on YOLO-barQR algorithm

Technical Field

The invention relates to the field of computer software, in particular to a multi-code quick reading method based on a YOLO-barQR algorithm.

Background

At present, code scanning application in the market is mainly performed on a single code, for example, a code scanning device shoots an image containing a mobile phone payment code, then decodes the payment code in the image, and sends a decoding result to a cloud to complete payment. With the popularization of code scanning application, some industries need to process a plurality of codes in the same image, for example, in the lottery industry, a plurality of different bar codes are needed on one lottery for security verification; for example, a large amount of information is required by some special industries to be carried by two-dimensional codes, the traditional single code cannot meet the requirement, and the information carrying of the two-dimensional codes is required to be expanded by including a plurality of two-dimensional codes in the same image.

A conventional decoding module can decode only one code for one picture, and a conventional decoding algorithm needs to be modified if a plurality of codes are decoded, for example, chinese patent publication No. CN105787403B, which proposes: external conditions are required to ensure that the captured image contains at least one barcode (e.g., using a high-speed camera and limiting the user to have to place the barcode in a designated location), locating multiple codes in a picture requires scanning the entire picture multiple times, destroying the scanned codes in the image, consuming time, and easily damaging the image of the undeciphered codes; moreover, the positioning of the codes in the picture needs to be realized according to certain characteristic information of a certain code system, for example, the two-dimensional code QR needs to be positioned by arranging the two-dimensional code into an L shape according to 3 'return' fonts of the QR code, and the error rate of the mode is high, for example, when a plurality of QR codes are arranged in a close matrix type, the positioning and dislocation may cause failure of subsequent decoding, thereby missing the codes.

So, the sign indicating number in the location picture needs to be realized according to certain characteristic of certain code system, and different code systems need the different characteristics of special study, wherein has the problem in two aspects: on one hand, the characteristics of some two-dimensional codes are obvious and have huge differences, and after the characteristics are specially researched, special software needs to be written for realization, so that the research and development workload and the research and development period are huge; on the other hand, the characteristics of some one-dimensional codes are not obvious, the difference between different one-dimensional codes is very small, and effective characteristics cannot be researched at all, so that the error positioning or the missing positioning is caused. Certain feature information of a certain code system needs to be analyzed in a high-pixel image, and if the pixel is too low, certain feature positioning can be failed.

Accordingly, the prior art is deficient and needs improvement.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the multi-code rapid recognizing and reading method based on the YOLO-barQR algorithm can recognize all codes in a picture at one time, accurately position each code, does not depend on the resolution of the picture, does not need to research the characteristics of each code system, and saves a large amount of research and development investment.

The technical scheme of the invention is as follows: a multi-code fast reading method based on a YOLO-barQR algorithm comprises the following steps: step S1: acquiring a picture of a graph to be identified; step S2: adopting a YOLO-barQR network model after the data set picture training to identify all codes in the picture; step S3: adopting the YOLO-barQR network model to position the coordinates of each code and evaluating the quality weight of each code; step S4: and sorting according to the quality weight, sequentially cutting the graph of each code according to the coordinates to obtain pictures, and outputting the pictures to a decoding module for decoding.

By applying the technical scheme, in the multi-code rapid reading method, in step S4, pictures are sequentially cut out of the graph of each code according to the coordinates according to the quality weight from high to low, and the pictures are sequentially output to a decoding module for decoding.

By applying each technical scheme, in the multi-code rapid reading method, before the step S2, the YOLO-barQR network model is trained, wherein during the training, data enhancement including exposure, saturation and blur is performed on the pictures of the data set, and each code pattern obtained through training and recognition is converted into a picture suitable for the YOLO-barQR network model.

By applying each technical scheme, in the multi-code rapid identification and reading method, when the YOLO-barQR network model is trained, the mean square error is adopted as a loss function.

By applying each technical scheme, in the multi-code rapid reading method, the structure of the YOLO-barQR network model is as follows:

serial number

Type (B)

Filter with a filter element having a plurality of filter elements

Size of

Input device

Output of

Activating a function

1

Convolutional layer

16

3x3/1

288×288×1

288×288×16

ReLU

2

Pooling layer

2x2/2

288×288×16

144×144×16

ReLU

3

Convolutional layer

32

3x3/1

144×144×16

144×144×32

ReLU

4

Pooling layer

2x2/2

144×144×32

72×72×32

ReLU

5

Convolutional layer

32

3x3/1

72×72×32

ReLU

6

Pooling layer

2x2/2

72×72×32

36×36×32

ReLU

7

Convolutional layer

64

3x3/1

36×36×32

36×36×64

ReLU

8

Pooling layer

2x2/2

36×36×64

18×18×64

ReLU

9

Convolutional layer

128

3x3/1

18×18×64

18×18×128

ReLU

10

Pooling layer

2x2/2

18×18×128

9×9×128

ReLU

11

Convolutional layer

256

3x3/1

9×9×128

9×9×256

ReLU

12

Convolutional layer

35

1x1/1

9×9×256

9×9×35

Liner

By applying the technical scheme, in the multi-code quick reading method, after the step S4, the picture is shot and stored, and the corresponding code is drawn in the picture according to the coordinate of each code.

By applying the technical scheme, in the multi-code rapid reading method, the graphs to be identified in the picture are one-dimensional codes, two-dimensional codes and postal codes.

By adopting the scheme, the YOLO-barQR network model trained by the data set picture is applied to the identification field of various codes, all the codes in the picture can be identified at one time based on a YOLO-barQR algorithm, and a plurality of codes with the same code system and compact arrangement in the picture can be accurately positioned, so that the accuracy rate is higher; and a plurality of codes of different code systems in the picture can be accurately positioned, the characteristics of each code system do not need to be researched, and a large amount of research and development investment is saved. Based on a YOLO-barQR algorithm, the code can be accurately positioned by the ordinary resolution without depending on the resolution of the image. And, the decoding order may adopt a parallel decoding, a sequential decoding or a random decoding manner. The quality weight of each code can be evaluated and then decoded in the quality weight order, and the first decoding result can be output more quickly.

Drawings

FIG. 1 is a flow chart of an embodiment of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and the specific embodiments.

The embodiment provides a multi-code fast reading method based on a YOLO-barQR algorithm, as shown in fig. 1, the multi-code fast reading method includes the following steps: first, step S1: acquiring a picture of a graph to be identified; then by step S2: adopting a YOLO-barQR network model after the data set picture training to identify all codes in the picture, scanning the picture puzzle to obtain an image when obtaining an identification graph, and judging whether a bar code exists in the picture; then, in step S3: adopting the YOLO-barQR network model to position the coordinates of each code and evaluating the quality weight of each code; step S4: the images of each code are sequentially cut according to the coordinates according to the quality weight sequence and then output to a decoding module for decoding, wherein the images of each code are sequentially cut according to the coordinates according to the quality weight from high to low and then output to the decoding module for decoding; and, the decoding module is an existing conventional decoding module.

Before the step S2, the YOLO-barQR network model is trained, wherein during training, data enhancement including exposure, saturation, and blur is performed on the picture of the data set, and each code pattern obtained by training and recognition is converted into a picture suitable for the YOLO-barQR network model, that is, during training, various data enhancement including exposure, saturation, blur, and the like is performed on the data set, so that diversity of the sample can be enriched, and during recognition, the dimension code picture obtained by relatively stable scanning is normalized into a picture suitable for the YOLO-barQR network, so as to improve the recognition rate and the recognition speed.

Wherein, the structure of the YOLO-barQR network model is as follows:

as shown in the above table, the structure of the YOLO-barQR network model includes a feature extraction network and a position/category detection network, the feature extraction network is configured as a combination of a plurality of convolution layers and pooling layers, each convolution layer includes a plurality of convolution kernel filters for extracting features of each input code, and outputs a feature extraction map; the position type detection network comprises a convolutional layer, wherein the convolutional layer comprises a plurality of convolutional kernel filters for detecting the characteristics of the characteristic extraction graph and obtaining the position and the type of an input code.

Wherein, the code that YOLO-barQR network model discerned includes two-dimensional code, one-dimensional code and zip code, and wherein, the two-dimensional code includes: QR Code, Micro QR, DataMatrix, PDF417, MicroPDF417, Codablock A/F, Aztec Code, Maxicode, Han Xin, etc.; the one-dimensional code includes: EAN-13 (with an additional Code), EAN-8 (with an additional Code), UPC-A (with an additional Code), UPC-E (with an additional Code), ISSN, ISBN, Codabar, Code 128, GS1-128, Code 39 Full ASCII, Code 93, ITF-16, ITF-14, Interleaved 2 of5, Industrial 2 of5, Matrix 2 of5, Standard 2 of5, NEC 2 of 2, GS1 Databar, GS1 Databar extended, GS1 Databar Limited, GS1 Composite Code, Code 11, Telepen, MSI-Plessey, Plessey, Pharmacode, etc.; the postal code includes: chian Post, Korea Post, Japan Post, PLANET, POSTNET, KIXCode, and the like.

And when the YOLO-barQR network model is trained, the data set is subjected to various data enhancements including exposure, saturation, blur and the like, so that the diversity of the sample is enriched

The dimension clustering method of the YOLO-barQR network model is that target frames marked manually in a data set are clustered through k-means, the statistical rule of the target frames is found, the number k of clusters is used as the number of anchors, and the width dimension and the height dimension of k cluster centers box are used as the dimension of the anchors. And carrying out cluster analysis on target frames corresponding to the target areas in the self-collected data set to obtain the optimal number of anchors and width and height dimensions suitable for detecting the data set.

And, the selection of the loss function of the YOLO-barQR network model is: the loss function of YOLOv2 is the same as YOLOv1, and for a group channel in the training set, which cell the center falls in, then the bounding box corresponding to 5 anchors of the cell is responsible for predicting it, and specifically, which prediction is also determined according to the calculated card threshold of the IOU, and the one with the largest IOU value is selected finally. This is also true in the case where each Cell contains at most one target, and in fact, there are substantially no more than 1. The prior box on the match with the ground truth is responsible for calculating the coordinate error, confidence error and classification error, while the other 4 bounding boxes only calculate the confidence error.

The loss function can be represented by the following graph:

in calculating the [ formula ] and [ formula ] errors for boxes, the square root is used in YOLOv1 to reduce the effect of box size on the error, while YOLOv2 is a direct calculation, but the weight coefficients are modified according to the size of the ground channel: (2-treth. w) treth. h) (where [ formula ] and [ formula ] are both normalized to (0,1)), the weighting factor for the smaller scale [ formula ] will be larger, amplifying the error, and achieving a similar effect as the square root calculation of YOLOv 1.

The loss function expression of the YOLO-barQR network model is as follows:

yolov3 Loss is the sum of three feature maps Loss:

Loss＝LossN1+LossN2+LossN3

and lambda is a weight constant, the proportion among the detection frames Loss, the obj confidence Loss and the nonobj confidence Loss is controlled, the number of negative examples is more than dozens times of that of positive examples, and the detection effect can be controlled through the weight super-parameter.

If the result is positive, 1 is output, and otherwise, 0 is output;

if the negative example is true, 1 is output, otherwise, 0 is output; ignoring the samples all outputs 0.

x, y, w, h use MSE as a loss function and also smooth L1 loss (ex Faster R-CNN) as a loss function. smooth L1 may make training smoother. Confidence, class label, is a 0,1 binary classification, so cross entropy is used as a loss function.

After step S4, the picture is captured and stored, and the corresponding code is outlined in the picture according to the coordinates of each code. The actual operation result shows that the system can rapidly mark each dimension code within 6ms under the calculation capability of two calculation forces, obtain the upper left coordinate, the upper right coordinate and the lower left coordinate of each dimension code and mark the category.

Thus, in the embodiment, through the YOLO-barQR algorithm of the YOLO-barQR network model, whether the barcode exists in the current image can be quickly identified without external limiting conditions, and the next decoding is carried out only when the barcode exists; moreover, all the bar codes in the image can be identified at one time; in addition, the method scans the picture once, positions the coordinates of all the bar codes through a YOLO-barQR algorithm, cuts out a part of the picture according to the coordinates and sends the part of the picture to a next decoding module, and the un-decoded picture cannot be damaged.

The method can accurately position a plurality of codes with the same code system and close arrangement in the picture based on the YOLO-barQR algorithm, and has higher accuracy.

The method is based on the YOLO-barQR algorithm, can accurately position a plurality of codes of different code systems in the picture under the condition that a large number of various codes are trained, does not need to research the characteristics of each code system, and saves a large amount of research and development investment.

The method adopts a YOLO-barQR algorithm, does not depend on the resolution of an image, and can accurately position the bar code with ordinary resolution.

The method can evaluate the quality weight of each code through a YOLO-barQR algorithm, then carries out decoding according to the quality weight sequence, and can output the first decoding result more quickly.

The method can automatically identify whether the bar code exists in the image, so the equipment using the method can work in an induction mode, namely, the equipment can immediately decode when the bar code exists, and does not need a user to press a key to photograph for decoding, thereby bringing higher working efficiency to the user.

According to the method, the bar code can be drawn and framed according to the identified bar code coordinate in the shot and stored picture, so that the user experience is improved.

The present invention is not limited to the above preferred embodiments, and any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-code fast reading method based on a YOLO-barQR algorithm is characterized by comprising the following steps: the method comprises the following steps:

step S1: acquiring a picture of a graph to be identified;

step S2: adopting a YOLO-barQR network model after the data set picture training to identify all codes in the picture;

step S3: adopting the YOLO-barQR network model to position the coordinates of each code and evaluating the quality weight of each code;

step S4: and sorting according to the quality weight, sequentially cutting the graph of each code according to the coordinates to obtain pictures, and outputting the pictures to a decoding module for decoding.

2. The multi-code fast reading method according to claim 1, characterized in that: in step S4, the pictures are sequentially cut out from the graph of each code according to the coordinates according to the quality weight from high to low, and the pictures are sequentially output to the decoding module for decoding.

3. The multi-code fast reading method according to claim 1, characterized in that: before the step S2, the YOLO-barQR network model is also trained, wherein during the training, the pictures of the data set are subjected to data enhancement including exposure, saturation, and blur, and each code pattern obtained through the training and recognition is also converted into a picture suitable for the YOLO-barQR network model.

4. The multi-code fast reading method according to claim 3, wherein: in training the YOLO-barQR network model, the mean square error is used as a loss function.

5. The multi-code fast reading method according to claim 3, wherein: the structure of the YOLO-barQR network model is as follows:

6. the multi-code fast reading method according to claim 1, characterized in that: and after the step S4, shooting and keeping the picture, and drawing a line and framing the corresponding code in the picture according to the coordinate of each code.

7. The multi-code fast reading method according to any one of claims 1 to 6, wherein: the graphs to be identified in the picture are one-dimensional codes, two-dimensional codes and postal codes.