阅读说明：本技术 基于rfid和深度学习的兽药生产指针式仪表图像识别方法和装置 (Animal remedy production pointer instrument image identification method and device based on RFID and deep learning ) 是由 曹姗姗 孙伟 孔繁涛 吴建寨 金忠明 邱琴 于 2019-11-25 设计创作，主要内容包括：本发明提供了一种基于RFID和深度学习的兽药生产指针式仪表图像识别方法,包括以下步骤：S1、采集仪表图像；S2、提取表盘区域；S3、图像预处理；S4、指针定位拟合；S5、仪表示数判定；S6、仪表预警判断。本发明提供的识别方法可以存储大量的电子数据,解决了工作人员在兽药生产车间有毒有害、有辐射、易污染的环境下所带来的健康损伤难题,最大程度保障工作人员的人身安全,还具有读数快、准确度高、适应性强、管理智能化的优点,可对异常环境及时进行预警功能,极大提高了兽药生产的效率,保障了兽药存储环境和生产环境的安全性。(The invention provides a veterinary drug production pointer instrument image identification method based on RFID and deep learning, which comprises the following steps: s1, collecting an instrument image; s2, extracting a dial area; s3, preprocessing the image; s4, positioning and fitting the pointer; s5, judging the number of the instrument; and S6, early warning and judgment of the instrument. The identification method provided by the invention can store a large amount of electronic data, solves the problem of health damage of workers in toxic, harmful, radiant and easily-polluted environments in an animal medicine production workshop, ensures the personal safety of the workers to the maximum extent, has the advantages of quick reading, high accuracy, strong adaptability and intelligent management, can perform an early warning function on abnormal environments in time, greatly improves the efficiency of animal medicine production, and ensures the safety of animal medicine storage environments and production environments.)

1. A veterinary drug production pointer instrument image identification method based on RFID and deep learning is characterized by comprising the following steps:

s1, collecting instrument data: acquiring an image of a pointer instrument dial plate containing a reading to be identified in a veterinary drug production workshop;

s2, extracting dial areas: detecting the image acquired in S1 by using a fast-RCNN algorithm trained based on deep learning, rapidly positioning the coordinates of the dial area, and extracting the dial image;

s3, image preprocessing: graying, image filtering and denoising and binarization preprocessing are carried out on the dial plate image extracted in the S2, and a dashboard binarization image only containing black and white pixels is obtained;

s4, pointer positioning fitting: performing final pointer positioning and fitting on the preprocessed instrument panel image by adopting a probabilistic Hough transform algorithm to obtain coordinate information and length information of a pointer;

s5, judging the meter number: performing digital identification based on a morphological operation method and combined with an Le Net-5 convolutional neural network, and performing reading judgment on a pointer instrument by using a distance method by using the coordinate information and the length information of the instrument panel pointer extracted in S4;

s6, early warning judgment of the instrument: and writing the meter reading judged in the step S5 into the RFID electronic tag 6 database, judging whether the early warning threshold value in the database is reached, and performing corresponding early warning operation.

2. The animal remedy production pointer instrument image identification method based on RFID and deep learning of claim 1, wherein the fast-RCNN algorithm in S2 comprises the following steps:

s201, RPN extracts a candidate region;

s202, extracting features by a deep learning network (CNN);

s203, Softmax classification;

and S204, performing multi-task loss function frame regression.

3. The animal remedy production pointer instrument image identification method based on RFID and deep learning as claimed in claim 1, wherein the image preprocessing in S3 comprises the following steps:

s301, graying processing: converting the color image into a gray image by adopting a three-component weighted average method, wherein the calculation formula is as follows:

Gray(i,j)＝0.299R(i,j)+0.578G(i,j)+0.114B(i,j)

wherein: gray (i, j) represents the pixel value of a pixel point of a Gray picture, R represents red space pixel numerical value information, G represents green space pixel numerical value information, and B represents blue space pixel numerical value information;

s302, image filtering and denoising: removing Gaussian noise in an original image by adopting a Gaussian filtering algorithm, wherein the Gaussian filtering is realized in a form of a template, namely a Gaussian kernel, the Gaussian kernel is obtained by discretizing a two-dimensional Gaussian function, and the two-dimensional Gaussian function formula is as follows:

wherein: sigma is a standard deviation, Gaussian kernels with different standard deviations are different, the weighting coefficient of the Gaussian kernels accords with Gaussian distribution, the weight of the central point is higher, the weight is far away from the central point, and the weight is rapidly attenuated;

s303, binarization preprocessing: the original image with 256 gray levels is converted into an image consisting of black and white pigments by adopting a maximum inter-class variance method, and the variance is used for representing the difference between classes.

4. The animal remedy production pointer instrument image identification method based on RFID and deep learning of claim 1, wherein the probabilistic Hough transform algorithm in S4 is: only partial characteristic pixel points are operated in the original binary image, and the selection of the characteristic pixel points has randomness and is used for detecting two end points of a pointer in the image.

5. An animal remedy production pointer instrument image recognition device based on RFID and deep learning is characterized by comprising an image acquisition module (1), a knowledge base module (2), an image recognition module (3) and an early warning module (4);

the image acquisition module (1) comprises a pointer instrument, an RFID electronic tag (5), an RFID reader-writer (6) and a camera unit (7), wherein the RFID electronic tag (5) used for storing relevant data of the pointer instrument is attached to the position right below the pointer instrument, the camera unit (7) is installed right in front of the pointer instrument, and the RFID reader-writer (6) is a handheld reading-writing device and is connected with the RFID electronic tag (5);

the early warning module (4) comprises a display screen (9) for displaying alarm characters, an audible and visual alarm (8) and a communication component (10) for sending alarm information to workers;

the knowledge base module (2) is a database for storing relevant data of the instrument;

the image recognition module (1) is a computer used for recognizing and reading an instrument image and transmitting a reading result to the knowledge base module (2) and the early warning module (4); the image acquisition module (1) is in wireless connection with the image recognition module (3), and the knowledge base module (2) is simultaneously connected with the image acquisition module (1), the image recognition module (3) and the early warning module (4).

6. The animal medicine production pointer instrument image recognition device based on RFID and deep learning of claim 5, wherein the camera unit (7) comprises a camera body (11), a camera (24) and an incandescent lamp (18); the camera body (11) is fixed through the camera fixing seat (12), the camera body (11) is connected with the camera (24), the camera (24) is covered with the dust cover (25), and the camera body (11) is rotatably connected with the incandescent lamp (18).

Technical Field

The invention belongs to the technical field of instrument data identification, and particularly relates to a pointer instrument image identification method and device for veterinary drug production based on RFID and deep learning.

Background

With the progress of science and technology, veterinary biological products and production modes are continuously innovated, but still many problems exist, most of the existing reading of instrument data of veterinary production workshops adopts wired remote transmission or manual reading, and the traditional modes have large workload, low efficiency and easy error; meanwhile, the veterinary drug is extremely sensitive to the environmental requirements, so that the precise management of the environments such as temperature, humidity and the like is realized, and the suitability of the drug storage conditions is ensured; meanwhile, toxic gas or radiation is easily generated during production, so that the health of workers is greatly damaged, and the danger existing in a production workshop is effectively monitored in time. Therefore, a method and a device for intelligently reading meters in an animal medicine production workshop are needed.

Disclosure of Invention

The invention aims to solve the technical problems that the image recognition method and the image recognition device for the animal remedy production pointer instrument based on the RFID and the deep learning are provided aiming at the defects of the prior art, the device is scientific and simple in structure, strong in practicability and low in cost, can quickly and automatically read the instrument information of an animal remedy production workshop, and has the characteristics of contact-free operation, long reading distance, good penetrability, long service life, strong pollution resistance, strong durability and the like.

In order to solve the technical problems, the invention adopts the technical scheme that: a veterinary drug production pointer instrument image identification method based on RFID and deep learning is characterized by comprising the following steps:

s1, collecting meter data, collecting images of a pointer type meter dial plate containing a reading to be identified in the veterinary drug production workshop by using a high-definition camera, and acquiring meter related data stored in the RFID electronic tag;

s2, extracting a dial area, detecting the image acquired in S1 by using a fast-RCNN algorithm trained based on deep learning, rapidly positioning the coordinates of the dial area, and extracting the area as a dial image;

s3, image preprocessing, namely performing graying, image filtering and denoising and binarization preprocessing on the dial plate image extracted in the S2 to obtain a binarized image of the instrument panel only containing black and white pixels;

s4, pointer positioning fitting, namely performing final pointer positioning and fitting on the preprocessed dial plate image by adopting an improved probability Hough transform algorithm to obtain coordinate information and length information of the pointer;

s5, meter registration judgment, namely performing digital identification based on a morphological operation method and in combination with an Le Net-5 convolutional neural network, and performing registration judgment on a pointer instrument by using a distance method by using extracted instrument panel pointer information and numerical value information;

s6, performing early warning judgment on the instrument, automatically writing the finally obtained dial plate indicating result into the RFID electronic tag 6 database, judging whether the dial plate indicating result reaches an early warning threshold value in the database, and ensuring the safety of the veterinary drug production workshop.

Preferably, the fast-RCNN algorithm comprises the steps of:

s201, RPN extraction candidate area: the RPN is realized by extracting n-dimensional feature vectors through sliding windows of feature maps output after sharing convolutional layers, K possible candidate regions are generated for any point on the feature maps, foreground and background probabilities of the candidate regions and prediction of a real candidate frame are carried out through two layers of full connection, and the candidate frames output by the RPN are classified and abandoned through an IOU.

The detection area is extracted by adopting a full convolution network, the boundary and the fraction of a target are predicted by performing end-to-end training of reverse propagation and random gradient reduction on a task for generating a detection candidate frame, the convolution characteristic of a full image is shared with a subsequent detection network, and the speed of area candidate is accelerated.

S202, deep learning network CNN extraction features: adopting a region downsampling mode, and adopting corresponding downsampling kernels aiming at different candidate regions, thereby ensuring that the input feature graphs of the Fast-RCNN network have the same size; and simultaneously, performing further feature extraction on the feature map with the same scale through the full connection layer and the Drop Out layer.

S203, Softmax classification: softmax is used as the extension of logistic regression, so that the algorithm can be applied to the classification problem of multiple classes; and (4) further extracting the features of the feature maps with the same scale, and finally sending the feature maps into a Softmax layer for classification and regression.

S204, multi-task loss function frame regression: the multi-task loss function directly adds the frame regression to the CNN network for training, improves the classification of the original loss function only aiming at the identification category, and measures the frame regression by adopting the loss function, thereby adding the frame regression to the training of the network.

Preferably, the image preprocessing described in S3 includes the steps of:

s301, graying, namely converting the color image into a grayscale image by adopting a three-component weighted average method, wherein the calculation formula is as follows:

Gray(i,j)＝0.299R(i,j)+0.578G(i,j)+0.114B(i,j)

wherein: gray (i, j) represents the pixel value of a pixel point of a grayed picture, R represents the pixel value information of a red space, G represents the pixel value information of a green space, and B represents the pixel value information of a blue space.

S302, filtering and denoising the image, and removing Gaussian noise in the original image by adopting a Gaussian filtering algorithm to achieve the purpose of smoothing the image. The Gaussian filtering is realized in a mode of a template, namely a Gaussian kernel, the Gaussian kernel is obtained by discretizing a two-dimensional Gaussian function, and the two-dimensional Gaussian function formula is as follows:

wherein: sigma is standard deviation, and Gaussian kernels with different standard deviations are different. The weighting coefficient of the Gaussian kernel conforms to Gaussian distribution, the weight of the central point is higher, and the weight is rapidly attenuated when the central point is far away from the central point.

S303, binarization preprocessing, namely converting the original image with 256 gray levels into an image consisting of black and white pigments by adopting a maximum inter-class variance method (OTSU algorithm), representing the difference between classes by using variance, and strengthening the edge of the image and highlighting the characteristics of the image. The detection of the line segments of the binary image can be effectively realized through a probabilistic Hough transform algorithm; the used threshold value is fixed, so that a plurality of features can be fitted in the binarized instrument panel image, namely a plurality of line segments exist after the binarized instrument panel image is detected by a probability Hough transform algorithm; and calculating the length of the line segment by using the end point coordinates of the plurality of line segments obtained by the probabilistic Hough transform algorithm, and selecting the line segment with the maximum value according to the length of the line segment for fitting the pointer of the instrument panel, thereby completing the positioning and fitting of the pointer of the instrument panel.

Preferably, the probabilistic hough transform algorithm in S4 is: only partial characteristic pixel points are operated in the original binary image, the selection of the characteristic pixel points has certain randomness, and two end points of a pointer in the image can be detected, so that the position of the pointer in the image is accurately positioned.

Preferably, the morphological operation is: firstly, the corrosion operation is used for removing noise points, and then the expansion operation is used for enhancing the digital image.

Preferably, the Le Net-5 convolutional neural network adopts two convolutional layers, two sampling layers and two full-connection layers, and the activation function adopts a Relu function; network training is completed under a training set, wherein the training set comprises 60000 data sets, 50000 sample images are used for training, an optimization mode adopted in the training process is a momentum descent method, and 10000 sample images are left for testing after the training is completed.

In addition, the invention also provides an animal remedy production pointer instrument image recognition device based on RFID and deep learning, which is characterized by comprising an image acquisition module, a knowledge base module, an image recognition module and an early warning module; the image acquisition module is used for acquiring pointer instrument images and acquiring instrument related data stored in the knowledge base module, and meanwhile, the data output end is in wireless connection with the image identification module through the wireless transmission module; the image recognition module is used for recognizing and reading an instrument image, and transmitting a reading result to the knowledge base module and the early warning module, and the early warning module judges the recognition result and an early warning threshold value; if the early warning threshold value is exceeded, an early warning event is immediately generated; if not, the instrument pointer does not enter the early warning area, and an early warning event does not need to be generated.

The image acquisition module comprises a pointer instrument, an RFID electronic tag, an RFID reader-writer and a camera unit, wherein the RFID electronic tag used for storing relevant data of the pointer instrument is attached to the position under the pointer instrument, the camera unit is installed right in front of the pointer instrument, and the RFID reader-writer is a handheld reading-writing device and is connected with the RFID electronic tag and used for reading, writing and recording information of the pointer instrument on the RFID electronic tag.

The early warning module comprises a display screen for displaying warning characters, an audible and visual alarm and a communication component for sending warning information to workers.

The knowledge base module is a database for storing relevant data of the instrument.

The image identification module is a computer used for identifying and reading an instrument image and transmitting a reading result to the knowledge base module and the early warning module; the image acquisition module is in wireless connection with the image identification module.

Preferably, the image recognition module includes: a processor, an input device, a communication interface, an output device, and a memory. The processor processes the collected information and sends out an instruction; the input equipment is used for inputting data and information and comprises a mouse, a keyboard and a high-definition camera; the communication interface is used for the communication function between the systems, including wireless communication interface and system communication interface; the wireless communication interface is in wireless connection with the RFID reader-writer by adopting a WIFI wireless communication interface circuit and is used for receiving RFID electronic tag information transmitted by the RFID operation state identification and wireless transmission equipment; the system communication interface adopts a USB or WIFI system communication interface for data transmission with background system software; the output device is a display for outputting data and information; the memory is used for storing programs.

Preferably, the camera unit includes a meter fixing base, a meter stand, a meter panel, a camera fixing base, a camera body, a rotating motor, a rotating base, and an incandescent lamp; the instrument fixing seat is placed on the workbench and connected with the instrument panel through the instrument support; paste under the panel board and have RFID electronic tags, inside instrument pointer and the instrument scale mark of setting up of panel board, the camera fixing base is connected with the camera body, sets up protection casing and lighting fixture on the camera body and erects the roof beam, the lighting fixture erects the roof beam and is connected with the rotating electrical machines, the roating seat is installed inside the rotating electrical machines, and the incandescent lamp passes through the bulb connector and the lighting fixture crossbeam links to each other with the roating seat, and the protection casing is used for avoiding the camera to receive influences such as dust, debris, and rotating electrical machines and roating seat are used for adjusting the bulb position, and the incandescent lamp is used for reducing the influence that external illumination changes and brings, guarantees illuminance and quality.

Compared with the prior art, the invention has the following advantages:

1. the identification device provided by the invention has the advantages of scientific and simple structure, strong practicability and low cost, the provided identification method can quickly and automatically read the instrument information of the veterinary drug production workshop, has the characteristics of contact-free operation, long reading distance, good penetrability, long service life, strong pollution resistance and durability and the like, realizes the automatic identification and reading tasks of the pointer instrument in the complex environment of the veterinary drug production workshop, improves the accuracy and stability, and can meet the actual application requirements of the veterinary drug production workshop.

2. The invention is based on a deep learning high-efficiency recognition algorithm, adopts fast-RCNN to detect and position the dial area, and combines a LeNet-5 method to carry out dial digital recognition, so that the system has the advantages of fast reading, high accuracy, strong adaptability and intelligent management.

3. Compared with a bar code technology, the RFID-based electronic tag has the advantages that the electronic tag can store a large amount of electronic data and can set password protection for reading and writing; meanwhile, the reader can dynamically read and identify the tag data in real time, in a contactless manner, in a remote manner and in batches by sending radio frequency signals through the antenna, so that the problem of health damage of workers in toxic, harmful, radiated and easily-polluted environments in a veterinary drug production workshop is solved while the workload is reduced to the maximum extent, and the personal safety of the workers is guaranteed to the maximum extent.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

Fig. 1 is an overall flowchart of the recognition method provided by the present invention.

Fig. 2 is a schematic view of the structure of the identification device of the present invention.

Fig. 3 is a schematic structural diagram of an image acquisition module in the present invention.

Fig. 4 is a schematic diagram of the composition structure of the RFID tag of the present invention.

Fig. 5 is a flow chart of the probabilistic hough transform algorithm of the present invention.

Description of reference numerals:

1-an image acquisition module; 2-knowledge base module; 3-an image recognition module;

4, an early warning module; 5-RFID electronic tags; 6-RFID reader;

7-a camera unit; 8, an audible and visual alarm; 9-a display screen;

10-a communication component; 11-camera body; 12-camera fixing base;

13-lamp holder vertical beam; 14-a rotating electrical machine; 15-a rotating base;

16-lamp holder cross beam; 17-bulb connector; 18-incandescent lamps;

19-instrument scale mark; 20-meter pointer; 21-instrument panel;

22-instrument holder; 23-instrument fixing seat; 24-a camera;

25-protective cover.

Detailed Description

As shown in fig. 1, the invention provides a pointer instrument image identification method for veterinary drug production based on RFID and deep learning, which comprises the following steps:

s1, collecting instrument images: a camera is used for collecting images of a pointer instrument dial plate containing a reading to be identified in the veterinary drug production workshop, and instrument related data stored in the RFID electronic tag 6 are obtained;

s2, extracting dial areas: detecting the image acquired in S1 by using a fast-RCNN algorithm trained based on deep learning, rapidly positioning the coordinates of a dial area, and extracting the area as a dial image;

s3, image preprocessing: graying, image filtering and denoising and binarization preprocessing are carried out on the dial plate image extracted in the S2, and a binarization image only containing black and white pixels of the instrument panel is obtained;

s4, pointer positioning fitting: performing final pointer positioning and fitting on the preprocessed dial plate image by adopting an improved probability Hough transform algorithm to obtain coordinate information and length information of the pointer;

s5, judging the meter number: performing digital identification based on a morphological operation method and combined with an Le Net-5 convolutional neural network, and performing reading judgment on a pointer instrument by using a distance method by using extracted instrument panel pointer information and numerical value information;

s6, early warning judgment of the instrument: and automatically writing the finally obtained dial plate number indicating result into the RFID electronic tag 6 database, and judging whether the early warning threshold value in the database is reached, so that the safety of the veterinary drug production workshop is guaranteed.

In this embodiment, the fast-RCNN algorithm includes the following steps:

s201, RPN extraction candidate area: the RPN network has the idea that n-dimensional feature vectors are extracted by sliding windows of feature maps output after the convolution layers are shared, K possible candidate regions are generated for any point on the feature maps, foreground and background probabilities of the candidate regions and the prediction of a real candidate frame are carried out through two layers of full connection, and the candidate frames output by the RPN network are classified and abandoned through an IOU (input output Unit) in a positioning precision evaluation mode.

Extracting a detection area by adopting a full convolution network, predicting the boundary and the fraction of a target by performing end-to-end training of reverse propagation and random gradient reduction on a task for generating a detection candidate frame, sharing the convolution characteristic of a full graph with a subsequent detection network, and accelerating the speed of area candidate;

s202, deep learning network CNN extraction features: adopting a region downsampling mode, and adopting corresponding downsampling kernels aiming at different candidate regions, thereby ensuring that the input feature graphs of the Fast-RCNN network have the same size; and simultaneously, performing further feature extraction on the feature map with the same scale through the full connection layer and the Drop Out layer.

S203, Softmax classification: softmax is used as the extension of logistic regression, so that the algorithm can be applied to the classification problem of multiple classes; and (4) further extracting the features of the feature maps with the same scale, and finally sending the feature maps into a Softmax layer for classification and regression.

S204, multi-task loss function frame regression: the multi-task loss function directly adds the frame regression to the CNN network for training, improves the classification of the original loss function only aiming at the identification category, and measures the frame regression by adopting the loss function, thereby adding the frame regression to the training of the network.

In this embodiment, the image preprocessing described in S3 includes the following steps:

s301, graying, namely converting the color image into a grayscale image by adopting a three-component weighted average method, wherein the calculation formula is as follows:

Gray(i,j)＝0.299R(i,j)+0.578G(i,j)+0.114B(i,j)

wherein: gray (i, j) represents the pixel value of a pixel point of a grayed picture, R represents the pixel value information of a red space, G represents the pixel value information of a green space, and B represents the pixel value information of a blue space.

S302, filtering and denoising the image, and removing Gaussian noise in the original image by adopting a Gaussian filtering algorithm to achieve the purpose of smoothing the image. The Gaussian filtering is realized in a mode of a template, namely a Gaussian kernel, the Gaussian kernel is obtained by discretizing a two-dimensional Gaussian function, and the two-dimensional Gaussian function formula is as follows:

wherein: sigma is standard deviation, and Gaussian kernels with different standard deviations are different. The weighting coefficient of the Gaussian kernel conforms to Gaussian distribution, the weight of the central point is higher, and the weight is rapidly attenuated when the central point is far away from the central point.

S303, binarization preprocessing, namely converting the original image with 256 gray levels into an image consisting of black and white pigments by adopting a maximum inter-class variance method (OTSU algorithm), representing the difference between classes by using variance, and strengthening the edge of the image and highlighting the characteristics of the image. The detection of the line segments of the binary image can be effectively realized through a probabilistic Hough transform algorithm; the used threshold value is fixed, so that a plurality of features can be fitted in the binarized instrument panel image, namely a plurality of line segments exist after the binarized instrument panel image is detected by a probability Hough transform algorithm; and calculating the length of the line segment by using the end point coordinates of a plurality of line segments obtained by probabilistic Hough transform, and selecting the line segment with the maximum value to fit the pointer of the instrument panel according to the length of the line segment, thereby completing the positioning and fitting of the pointer of the instrument panel.

As shown in fig. 5, in this embodiment, the probabilistic hough transform algorithm is: only partial characteristic pixel points are operated in the original binary image, the selection of the characteristic pixel points has certain randomness, and two end points of a pointer in the image can be detected, so that the position of the pointer in the image is accurately positioned. Specifically, the probabilistic hough transform firstly randomly extracts an edge point of the image, which is not on a straight line, and if the edge point does not exist, the algorithm is ended; then carrying out Hough transform on the point, and carrying out accumulation and calculation; and finally, selecting a point with the maximum value in the Hough space, if the point is larger than the threshold value, searching the linear section and outputting according to the relation with the threshold value, and if not, continuously searching the edge point. In the probability Hough transform algorithm, the used threshold value is fixed, so that a plurality of features can be fitted in the binarized instrument panel image, namely, a plurality of line segments exist in the binarized instrument panel image after the binarized instrument panel image is detected by the probability Hough transform algorithm; calculating the length of the line segment by utilizing the end point coordinates of a plurality of line segments obtained by probabilistic Hough transform, and then selecting the line segment with the maximum value to fit the pointer of the instrument panel according to the length of the line segment.

In this embodiment, the morphological operation method includes: firstly, the corrosion operation is used for removing noise points, and then the expansion operation is used for enhancing the digital image.

In this embodiment, the Le Net-5 convolutional neural network adopts two convolutional layers, two sampling layers and two full-connection layers, and the activation function adopts a Relu function; completing network training under a training set, wherein the training set comprises 60000 data sets; wherein 50000 sample images are used for training, the optimization mode adopted in the training process is momentum descent method, and 10000 sample images are used for testing after the training is finished.

As shown in fig. 2, the present embodiment further provides an animal remedy production pointer instrument image recognition device based on RFID and deep learning, which is characterized by comprising an image acquisition module 1, a knowledge base module 2, an image recognition module 3 and an early warning module 4. The image acquisition module 1 is used for acquiring pointer instrument images and acquiring instrument related data stored in the knowledge base module 2, and the data output end of the image acquisition module 1 is wirelessly connected with the image identification module 3 through a wireless transmission module; the image recognition module 3 is used for recognizing and reading an instrument image, and transmitting a reading result to the knowledge base module 2 and the early warning module 4, and the early warning module 4 is used for judging the recognition result and an early warning threshold value; if the early warning threshold value is exceeded, an early warning event is immediately generated; if not, the instrument pointer does not enter the early warning area, and an early warning event does not need to be generated.

As shown in fig. 3, the image acquisition module 1 includes a pointer instrument, an RFID electronic tag 5, an RFID reader/writer 6, and a camera unit, the RFID electronic tag 5 for storing data related to the pointer instrument is attached right below the pointer instrument, and the camera unit 7 is installed right in front of the pointer instrument; the pointer type instrument comprises an instrument fixing seat 23, an instrument support 22 and an instrument panel 21, and the camera unit 7 comprises a camera fixing seat 12, a camera body 11, a camera 24, a rotating motor 14, a rotating seat 15 and an incandescent lamp 18; the instrument fixing seat 23 is placed on the workbench and connected with the instrument panel 21 through the instrument support 22, the instrument pointer 20 and the instrument scale mark 19 are arranged inside the instrument panel 21, the camera fixing seat 12 is connected with the camera body 11, the camera body 11 is provided with the protective cover 25 for protecting the camera 24, the camera body 11 is fixedly connected with the lamp holder vertical beam 13, the lamp holder vertical beam 13 is provided with the rotating motor 14, the output shaft of the rotating motor 14 is fixedly connected with one end of the lamp holder cross beam 16 through the rotating seat 15, the other end of the lamp holder cross beam 16 is connected with the bulb connecting body 17 arranged on the incandescent lamp 18, the rotating motor 14 is used for adjusting the position of the incandescent lamp 18, the incandescent lamp 18 is used for reducing the influence caused by the change of external illumination and ensuring the illumination and the acquisition quality when an image is acquired, the RFID reader-writer 6 is a handheld read-write device and is, the pointer type meter reading and writing device is used for reading, writing and recording pointer type meter information on the RFID electronic tag 5.

The early warning module 4 comprises a display screen 9 for displaying alarm characters, an audible and visual alarm 8 and a communication component 10 for sending alarm information to workers.

The knowledge base module 2 is a database for storing relevant data of the instrument, and the knowledge base module 2 is connected with the image acquisition module 1, the image recognition module 3 and the early warning module 4 at the same time.

The image recognition module 3 is a computer for recognizing the instrument image, reading the instrument image and transmitting the reading result to the knowledge base module 2 and the early warning module 4; the image acquisition module 1 is in wireless connection with the image identification module 3.

In this embodiment, the image recognition module 3 includes: a processor, an input device, a communication interface, an output device, and a memory. The processor processes the collected information and sends out an instruction; the input equipment is used for inputting data and information and comprises a mouse, a keyboard and a high-definition camera; the communication interface is used for the communication function between the systems, including wireless communication interface and system communication interface; the wireless communication interface is in wireless connection with the REID reader-writer 6 by adopting a WIFI wireless communication interface circuit and is used for receiving RFID electronic tag 5 information transmitted by RFID operation state identification and wireless transmission equipment; the system communication interface adopts a USB or WIFI system communication interface for data transmission with background system software; the output device is a display for outputting data and information; the memory is used for storing programs.

As shown in fig. 4, in this embodiment, the RFID tag 5 includes: the device comprises a power supply, an antenna, a modulator, a code generator, a clock and a memory; the RFID tag 5 is mainly used to receive the radio frequency signal transmitted by the reader, and transmit the radio frequency signal to the tag chip in time for processing, and transmit the information of the tag chip to the RFID reader 6 by using the radio frequency signal. For passive tags, the antenna also needs to be responsible for providing energy for tag operation. The chip functions mainly to mediate and decode information received by the antenna, and at the same time, it is necessary to encode and modulate signals to be transmitted by the tag, execute an anti-collision algorithm, store data, and the like. Each tag has a unique electronic code associated with the target object. The program written inside the tag can be read and rewritten at any time according to the particular application. The signal to be transmitted by the RFID reader-writer 6 is loaded on a carrier signal with a certain frequency through encoding, the carrier signal is transmitted outwards through an antenna, and the RFID electronic tag 5 entering the working area of the RFID reader-writer 6 receives the pulse signal so as to obtain energy; a chip in the RFID electronic tag 5 modulates and demodulates signals, operates a read command or a write command, sends out self encoding information, is read and decoded by the RFID reader-writer 6, and then is transmitted to the image identification module 3 for relevant data processing; energy transfer, clock signal acquisition and data interaction exist between the RFID reader-writer 6 and the RFID electronic tag 5. The RFID electronic tag 5 obtains energy through a radio frequency field, the RFID reader-writer 6 transmits a clock signal to the RFID electronic tag 5, the RFID electronic tag 5 responds to a command of the RFID reader-writer 6 and transmits data to the RFID reader-writer 6 through a radio frequency module of the RFID electronic tag 5 or transmits the data to a memory in the RFID electronic tag 5 for storage.

The above description is only for the preferred embodiment of the present invention, and is not intended to limit the present invention in any way. Any simple modification, change and equivalent changes of the above embodiments according to the technical essence of the invention are still within the protection scope of the technical solution of the invention.