A Lightweight Neural Network-Based License Plate Recognition Method
A neural network and license plate recognition technology, applied in the direction of neural learning methods, biological neural network models, neural architectures, etc., can solve problems such as limitations of accuracy and generalization capabilities, speed impact, etc., to solve the misalignment of input and output, improve The effect of detecting speed and reducing the amount of parameters
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0064] The invention designs a license plate recognition method based on a lightweight neural network, adopts the CTC loss function to train the lightweight neural network model, realizes end-to-end training without segmentation, and effectively solves the situation that the input and output are not aligned, especially The following setting methods are adopted: including the preparation of the data set, the construction of the lightweight neural network model and the recognition of the license plate,
[0065] The preparation of the data forms a license plate picture with the license plate number as the file name, specifically:
[0066] First, obtain video screenshots from road videos, form sample images, and save them in the images folder; preferably, collect several videos of driving vehicles from the monitoring probe, and save one picture per second, a total of Collect more than 30,000 images (sample images) and save them to the images folder, where the sample images are all...
Embodiment 2
[0080] This embodiment is further optimized on the basis of the above-mentioned embodiment, and the same parts as the above-mentioned technical solutions will not be repeated here. In order to further realize the present invention, the following setting methods are specially adopted: the step 1) specifically It is: input the matrix with dimension [32, 3, 24, 92] into the first two-dimensional convolutional layer, after normalization operation (batch normalization) and nonlinear activation (relu), the dimension is [32, 64, 22 , 92] the first feature matrix x1; the first two-dimensional convolutional layer is provided with 64 convolution kernels with a stride of 1 and a size of 3×3.
Embodiment 3
[0082] This embodiment is further optimized on the basis of any of the above-mentioned embodiments, and the same parts as the foregoing technical solutions will not be repeated here. In order to better realize the present invention, the following setting method is specially adopted: the step 2 )Specifically:
[0083] 2.1) Input the first feature matrix x1 into the pooling layer to obtain a feature matrix of [32, 32, 20, 90]; in the step 2.1), the pooling area size is (1, 3, 3) and the step size is (1, 1, 1) pooling layer;
[0084] 2.2) Input the feature matrix obtained in step 2.1) into a special convolutional layer (named small_basic_block), and the output result after normalization operation (batch normalization) and nonlinear activation (relu), the dimension is [32, 128, 20, 90 ] of the second feature matrix x2.
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


