Seriously degraded license plate image information recovery method

A license plate image and degraded image technology, applied in the application field of image processing technology, can solve the problems of inability to deal with highly degraded images and incompleteness, and achieve the effect of great application value and reasonable design.

Inactive Publication Date: 2020-11-20
TIANJIN UNIVERSITY OF SCIENCE AND TECHNOLOGY
0 Cites 0 Cited by

AI-Extracted Technical Summary

Problems solved by technology

With the recent success of Convolutional Neural Networks (CNN), various overall image recognition methods have emerged. For example, Goodfellow et al. proposed a CNN architecture capable of recognizing arbitrary combinations of numbers of bounded length for extracting ...
View more

Abstract

The invention relates to a seriously degraded license plate image information recovery method. According to the method, a convolutional neural network is used to learn priori knowledge required by seriously degraded license plate information extraction; training and fine adjustment of the convolutional neural network are carried out by adopting the simulated license plate image and the real license plate image, and the character output probability of each license plate character position is obtained by adopting a SoftMax output form, so that a decision basis is provided for license plate information extraction, and unified processing of character segmentation and recognition is achieved. According to the method and the process, extraction of seriously degraded license plate information iseffectively realized, and extraction of all low-resolution license plate information is facilitated. The method can help the public security department to obtain meaningful evidences and clues, and has the maximum application value.

Application Domain

Character and pattern recognitionNeural architectures +1

Technology Topic

Computer visionInformation extraction +7

Image

  • Seriously degraded license plate image information recovery method

Examples

  • Experimental program(1)

Example Embodiment

[0022] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0023] Step 1. Collect various license plate images in the real environment;
[0024] Step 2. Analyze the characteristics of the license plate such as font, character distribution, color characteristics, etc., and design the license plate simulation generation program. Use the license plate simulation program to generate a large number of simulated license plate images.
[0025] Step 3. Add various degrees of noise information to the simulated license plate image and the real license plate image, and generate degraded images of different sizes accordingly. Then the size of these degraded images is enlarged to the same size, and the simulated degraded image database of license plate and the degraded image database of real license plate with label information are constructed. Construct training data set, verification data set and test data set of degraded images of simulated license plates; construct training data set, verification data set and test data sets of degraded images of real license plates;
[0026] Step 4. Construct a convolutional neural network and use the simulated license plate degradation image database for training;
[0027] Step 5. Use the real license plate degradation image database to fine-tune the model obtained in step 4 to further improve the model. Finally, the model needed to restore the degraded license plate image information is obtained.
[0028] Further, the license plate image in the step 1 is deformed, corrected, and cropped to become a rectangular license plate image containing only the license plate number (excluding the first Chinese character), and the size is 560*224 pixels;
[0029] Further, the license plate simulation generation program in the step 2 also simulates the brightness, color, and background noise of the license plate to form 1 million license plate images. The generated simulation image only contains the license plate number (excluding the first Chinese character). Rectangular license plate image, the size is 560*224 pixels;
[0030] Further, adding various degrees of noise information in step 3 is adding Gaussian noise of different programs, and the signal-to-noise ratio is expressed as: -3, 0, 3, 6, 18; the width (pixels) of images of different sizes is: 10, 16, 20, 28, 40, 56. The degraded image is uniformly enlarged to 140*56 pixels. The ratio of training data set, verification data set and test data set of simulated license plate degradation image is 7:2:1. The ratio of the training data set, the verification data set and the test data set of the real license plate degraded image is 7:2:1.
[0031] Further, the structure adopted by the convolutional neural network in the step 4: two layers of convolution plus one maximized pooling layer are a group, a total of 4 groups; then two layers of fully connected layers, the last fully connected layer and 7 Two output layers are connected. Each output layer has 36 units and performs SoftMax operation as the output layer output (domestic car license plates have 7 characters except the first Chinese character, and each character can take 26 letters and 10 numbers, a total of 36 Options). The cross-entropy loss function is used for network training, the batch size is 32, and the learning rate is 0.01.
[0032] Further, in the step 5, the model obtained in step 4 is fine-tuned, and the real license plate degraded image is used for training, and the learning rate is 0.0001.
[0033] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manners. Anyone skilled in the art is based on the technology of the present invention. Other implementations derived from the solution also fall within the protection scope of the present invention.

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.

Similar technology patents

Medical foldable bedside cabinet

Owner:苏州优点优唯医疗科技有限公司

Magnetic resonance imaging method of magnetization transfer and level choose inversion recovery combined pre-pulse

ActiveCN104856677AThe method steps are simplereasonable design
Owner:THE FIRST AFFILIATED HOSPITAL OF MEDICAL COLLEGE OF XIAN JIAOTONG UNIV

Karst area soil nutrient reserve sampling and calculating method

Owner:INST OF SUBTROPICAL AGRI CHINESE ACAD OF SCI

Classification and recommendation of technical efficacy words

  • high application value
  • reasonable design

Method for coal and gas outburst simulation experiment

Owner:CHINA COAL TECH ENG GRP CHONGQING RES INST

Embedding accuracy control construction method for embedded part of overlong equipment foundation platform

InactiveCN102134879Areasonable designSimple construction steps
Owner:SCEGC NO 6 CONSTR ENG GRP CO LTD

Frequency hopping communication anti-interference method based on cognitive radio

ActiveCN103441779AImprove anti-jamming communication abilityreasonable design
Owner:深圳市研谱科技有限公司
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products